net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32 #include <linux/siphash.h>
  33
  34 #include <linux/netfilter.h>
  35 #include <net/netlink.h>
  36 #include <net/sock.h>
  37 #include <net/netfilter/nf_conntrack.h>
  38 #include <net/netfilter/nf_conntrack_core.h>
  39 #include <net/netfilter/nf_conntrack_expect.h>
  40 #include <net/netfilter/nf_conntrack_helper.h>
  41 #include <net/netfilter/nf_conntrack_seqadj.h>
  42 #include <net/netfilter/nf_conntrack_l4proto.h>
  43 #include <net/netfilter/nf_conntrack_tuple.h>
  44 #include <net/netfilter/nf_conntrack_acct.h>
  45 #include <net/netfilter/nf_conntrack_zones.h>
  46 #include <net/netfilter/nf_conntrack_timestamp.h>
  47 #include <net/netfilter/nf_conntrack_labels.h>
  48 #include <net/netfilter/nf_conntrack_synproxy.h>
  49 #if IS_ENABLED(CONFIG_NF_NAT)
  50 #include <net/netfilter/nf_nat.h>
  51 #include <net/netfilter/nf_nat_helper.h>
  52 #endif
  53
  54 #include <linux/netfilter/nfnetlink.h>
  55 #include <linux/netfilter/nfnetlink_conntrack.h>
  56
  57 #include "nf_internals.h"
  58
  59 MODULE_LICENSE("GPL");
  60
  61 static int ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  62                                 const struct nf_conntrack_tuple *tuple,
  63                                 const struct nf_conntrack_l4proto *l4proto)
  64 {
  65         int ret = 0;
  66         struct nlattr *nest_parms;
  67
  68         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO);
  69         if (!nest_parms)
  70                 goto nla_put_failure;
  71         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  72                 goto nla_put_failure;
  73
  74         if (likely(l4proto->tuple_to_nlattr))
  75                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  76
  77         nla_nest_end(skb, nest_parms);
  78
  79         return ret;
  80
  81 nla_put_failure:
  82         return -1;
  83 }
  84
  85 static int ipv4_tuple_to_nlattr(struct sk_buff *skb,
  86                                 const struct nf_conntrack_tuple *tuple)
  87 {
  88         if (nla_put_in_addr(skb, CTA_IP_V4_SRC, tuple->src.u3.ip) ||
  89             nla_put_in_addr(skb, CTA_IP_V4_DST, tuple->dst.u3.ip))
  90                 return -EMSGSIZE;
  91         return 0;
  92 }
  93
  94 static int ipv6_tuple_to_nlattr(struct sk_buff *skb,
  95                                 const struct nf_conntrack_tuple *tuple)
  96 {
  97         if (nla_put_in6_addr(skb, CTA_IP_V6_SRC, &tuple->src.u3.in6) ||
  98             nla_put_in6_addr(skb, CTA_IP_V6_DST, &tuple->dst.u3.in6))
  99                 return -EMSGSIZE;
 100         return 0;
 101 }
 102
 103 static int ctnetlink_dump_tuples_ip(struct sk_buff *skb,
 104                                     const struct nf_conntrack_tuple *tuple)
 105 {
 106         int ret = 0;
 107         struct nlattr *nest_parms;
 108
 109         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP);
 110         if (!nest_parms)
 111                 goto nla_put_failure;
 112
 113         switch (tuple->src.l3num) {
 114         case NFPROTO_IPV4:
 115                 ret = ipv4_tuple_to_nlattr(skb, tuple);
 116                 break;
 117         case NFPROTO_IPV6:
 118                 ret = ipv6_tuple_to_nlattr(skb, tuple);
 119                 break;
 120         }
 121
 122         nla_nest_end(skb, nest_parms);
 123
 124         return ret;
 125
 126 nla_put_failure:
 127         return -1;
 128 }
 129
 130 static int ctnetlink_dump_tuples(struct sk_buff *skb,
 131                                  const struct nf_conntrack_tuple *tuple)
 132 {
 133         const struct nf_conntrack_l4proto *l4proto;
 134         int ret;
 135
 136         rcu_read_lock();
 137         ret = ctnetlink_dump_tuples_ip(skb, tuple);
 138
 139         if (ret >= 0) {
 140                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
 141                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 142         }
 143         rcu_read_unlock();
 144         return ret;
 145 }
 146
 147 static int ctnetlink_dump_zone_id(struct sk_buff *skb, int attrtype,
 148                                   const struct nf_conntrack_zone *zone, int dir)
 149 {
 150         if (zone->id == NF_CT_DEFAULT_ZONE_ID || zone->dir != dir)
 151                 return 0;
 152         if (nla_put_be16(skb, attrtype, htons(zone->id)))
 153                 goto nla_put_failure;
 154         return 0;
 155
 156 nla_put_failure:
 157         return -1;
 158 }
 159
 160 static int ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 161 {
 162         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 163                 goto nla_put_failure;
 164         return 0;
 165
 166 nla_put_failure:
 167         return -1;
 168 }
 169
 170 static int ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 171 {
 172         long timeout = nf_ct_expires(ct) / HZ;
 173
 174         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 175                 goto nla_put_failure;
 176         return 0;
 177
 178 nla_put_failure:
 179         return -1;
 180 }
 181
 182 static int ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 183 {
 184         const struct nf_conntrack_l4proto *l4proto;
 185         struct nlattr *nest_proto;
 186         int ret;
 187
 188         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 189         if (!l4proto->to_nlattr)
 190                 return 0;
 191
 192         nest_proto = nla_nest_start(skb, CTA_PROTOINFO);
 193         if (!nest_proto)
 194                 goto nla_put_failure;
 195
 196         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 197
 198         nla_nest_end(skb, nest_proto);
 199
 200         return ret;
 201
 202 nla_put_failure:
 203         return -1;
 204 }
 205
 206 static int ctnetlink_dump_helpinfo(struct sk_buff *skb,
 207                                    const struct nf_conn *ct)
 208 {
 209         struct nlattr *nest_helper;
 210         const struct nf_conn_help *help = nfct_help(ct);
 211         struct nf_conntrack_helper *helper;
 212
 213         if (!help)
 214                 return 0;
 215
 216         rcu_read_lock();
 217         helper = rcu_dereference(help->helper);
 218         if (!helper)
 219                 goto out;
 220
 221         nest_helper = nla_nest_start(skb, CTA_HELP);
 222         if (!nest_helper)
 223                 goto nla_put_failure;
 224         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 225                 goto nla_put_failure;
 226
 227         if (helper->to_nlattr)
 228                 helper->to_nlattr(skb, ct);
 229
 230         nla_nest_end(skb, nest_helper);
 231 out:
 232         rcu_read_unlock();
 233         return 0;
 234
 235 nla_put_failure:
 236         rcu_read_unlock();
 237         return -1;
 238 }
 239
 240 static int
 241 dump_counters(struct sk_buff *skb, struct nf_conn_acct *acct,
 242               enum ip_conntrack_dir dir, int type)
 243 {
 244         enum ctattr_type attr = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 245         struct nf_conn_counter *counter = acct->counter;
 246         struct nlattr *nest_count;
 247         u64 pkts, bytes;
 248
 249         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 250                 pkts = atomic64_xchg(&counter[dir].packets, 0);
 251                 bytes = atomic64_xchg(&counter[dir].bytes, 0);
 252         } else {
 253                 pkts = atomic64_read(&counter[dir].packets);
 254                 bytes = atomic64_read(&counter[dir].bytes);
 255         }
 256
 257         nest_count = nla_nest_start(skb, attr);
 258         if (!nest_count)
 259                 goto nla_put_failure;
 260
 261         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts),
 262                          CTA_COUNTERS_PAD) ||
 263             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes),
 264                          CTA_COUNTERS_PAD))
 265                 goto nla_put_failure;
 266
 267         nla_nest_end(skb, nest_count);
 268
 269         return 0;
 270
 271 nla_put_failure:
 272         return -1;
 273 }
 274
 275 static int
 276 ctnetlink_dump_acct(struct sk_buff *skb, const struct nf_conn *ct, int type)
 277 {
 278         struct nf_conn_acct *acct = nf_conn_acct_find(ct);
 279
 280         if (!acct)
 281                 return 0;
 282
 283         if (dump_counters(skb, acct, IP_CT_DIR_ORIGINAL, type) < 0)
 284                 return -1;
 285         if (dump_counters(skb, acct, IP_CT_DIR_REPLY, type) < 0)
 286                 return -1;
 287
 288         return 0;
 289 }
 290
 291 static int
 292 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 293 {
 294         struct nlattr *nest_count;
 295         const struct nf_conn_tstamp *tstamp;
 296
 297         tstamp = nf_conn_tstamp_find(ct);
 298         if (!tstamp)
 299                 return 0;
 300
 301         nest_count = nla_nest_start(skb, CTA_TIMESTAMP);
 302         if (!nest_count)
 303                 goto nla_put_failure;
 304
 305         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start),
 306                          CTA_TIMESTAMP_PAD) ||
 307             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 308                                                cpu_to_be64(tstamp->stop),
 309                                                CTA_TIMESTAMP_PAD)))
 310                 goto nla_put_failure;
 311         nla_nest_end(skb, nest_count);
 312
 313         return 0;
 314
 315 nla_put_failure:
 316         return -1;
 317 }
 318
 319 #ifdef CONFIG_NF_CONNTRACK_MARK
 320 static int ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct,
 321                                bool dump)
 322 {
 323         u32 mark = READ_ONCE(ct->mark);
 324
 325         if (!mark && !dump)
 326                 return 0;
 327
 328         if (nla_put_be32(skb, CTA_MARK, htonl(mark)))
 329                 goto nla_put_failure;
 330         return 0;
 331
 332 nla_put_failure:
 333         return -1;
 334 }
 335 #else
 336 #define ctnetlink_dump_mark(a, b, c) (0)
 337 #endif
 338
 339 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 340 static int ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 341 {
 342         struct nlattr *nest_secctx;
 343         int len, ret;
 344         char *secctx;
 345
 346         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 347         if (ret)
 348                 return 0;
 349
 350         ret = -1;
 351         nest_secctx = nla_nest_start(skb, CTA_SECCTX);
 352         if (!nest_secctx)
 353                 goto nla_put_failure;
 354
 355         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 356                 goto nla_put_failure;
 357         nla_nest_end(skb, nest_secctx);
 358
 359         ret = 0;
 360 nla_put_failure:
 361         security_release_secctx(secctx, len);
 362         return ret;
 363 }
 364 #else
 365 #define ctnetlink_dump_secctx(a, b) (0)
 366 #endif
 367
 368 #ifdef CONFIG_NF_CONNTRACK_LABELS
 369 static inline int ctnetlink_label_size(const struct nf_conn *ct)
 370 {
 371         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 372
 373         if (!labels)
 374                 return 0;
 375         return nla_total_size(sizeof(labels->bits));
 376 }
 377
 378 static int
 379 ctnetlink_dump_labels(struct sk_buff *skb, const struct nf_conn *ct)
 380 {
 381         struct nf_conn_labels *labels = nf_ct_labels_find(ct);
 382         unsigned int i;
 383
 384         if (!labels)
 385                 return 0;
 386
 387         i = 0;
 388         do {
 389                 if (labels->bits[i] != 0)
 390                         return nla_put(skb, CTA_LABELS, sizeof(labels->bits),
 391                                        labels->bits);
 392                 i++;
 393         } while (i < ARRAY_SIZE(labels->bits));
 394
 395         return 0;
 396 }
 397 #else
 398 #define ctnetlink_dump_labels(a, b) (0)
 399 #define ctnetlink_label_size(a) (0)
 400 #endif
 401
 402 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 403
 404 static int ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 405 {
 406         struct nlattr *nest_parms;
 407
 408         if (!(ct->status & IPS_EXPECTED))
 409                 return 0;
 410
 411         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER);
 412         if (!nest_parms)
 413                 goto nla_put_failure;
 414         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 415                 goto nla_put_failure;
 416         nla_nest_end(skb, nest_parms);
 417
 418         return 0;
 419
 420 nla_put_failure:
 421         return -1;
 422 }
 423
 424 static int
 425 dump_ct_seq_adj(struct sk_buff *skb, const struct nf_ct_seqadj *seq, int type)
 426 {
 427         struct nlattr *nest_parms;
 428
 429         nest_parms = nla_nest_start(skb, type);
 430         if (!nest_parms)
 431                 goto nla_put_failure;
 432
 433         if (nla_put_be32(skb, CTA_SEQADJ_CORRECTION_POS,
 434                          htonl(seq->correction_pos)) ||
 435             nla_put_be32(skb, CTA_SEQADJ_OFFSET_BEFORE,
 436                          htonl(seq->offset_before)) ||
 437             nla_put_be32(skb, CTA_SEQADJ_OFFSET_AFTER,
 438                          htonl(seq->offset_after)))
 439                 goto nla_put_failure;
 440
 441         nla_nest_end(skb, nest_parms);
 442
 443         return 0;
 444
 445 nla_put_failure:
 446         return -1;
 447 }
 448
 449 static int ctnetlink_dump_ct_seq_adj(struct sk_buff *skb, struct nf_conn *ct)
 450 {
 451         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
 452         struct nf_ct_seqadj *seq;
 453
 454         if (!(ct->status & IPS_SEQ_ADJUST) || !seqadj)
 455                 return 0;
 456
 457         spin_lock_bh(&ct->lock);
 458         seq = &seqadj->seq[IP_CT_DIR_ORIGINAL];
 459         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_ORIG) == -1)
 460                 goto err;
 461
 462         seq = &seqadj->seq[IP_CT_DIR_REPLY];
 463         if (dump_ct_seq_adj(skb, seq, CTA_SEQ_ADJ_REPLY) == -1)
 464                 goto err;
 465
 466         spin_unlock_bh(&ct->lock);
 467         return 0;
 468 err:
 469         spin_unlock_bh(&ct->lock);
 470         return -1;
 471 }
 472
 473 static int ctnetlink_dump_ct_synproxy(struct sk_buff *skb, struct nf_conn *ct)
 474 {
 475         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
 476         struct nlattr *nest_parms;
 477
 478         if (!synproxy)
 479                 return 0;
 480
 481         nest_parms = nla_nest_start(skb, CTA_SYNPROXY);
 482         if (!nest_parms)
 483                 goto nla_put_failure;
 484
 485         if (nla_put_be32(skb, CTA_SYNPROXY_ISN, htonl(synproxy->isn)) ||
 486             nla_put_be32(skb, CTA_SYNPROXY_ITS, htonl(synproxy->its)) ||
 487             nla_put_be32(skb, CTA_SYNPROXY_TSOFF, htonl(synproxy->tsoff)))
 488                 goto nla_put_failure;
 489
 490         nla_nest_end(skb, nest_parms);
 491
 492         return 0;
 493
 494 nla_put_failure:
 495         return -1;
 496 }
 497
 498 static int ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 499 {
 500         __be32 id = (__force __be32)nf_ct_get_id(ct);
 501
 502         if (nla_put_be32(skb, CTA_ID, id))
 503                 goto nla_put_failure;
 504         return 0;
 505
 506 nla_put_failure:
 507         return -1;
 508 }
 509
 510 static int ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 511 {
 512         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 513                 goto nla_put_failure;
 514         return 0;
 515
 516 nla_put_failure:
 517         return -1;
 518 }
 519
 520 /* all these functions access ct->ext. Caller must either hold a reference
 521  * on ct or prevent its deletion by holding either the bucket spinlock or
 522  * pcpu dying list lock.
 523  */
 524 static int ctnetlink_dump_extinfo(struct sk_buff *skb,
 525                                   struct nf_conn *ct, u32 type)
 526 {
 527         if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 528             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 529             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 530             ctnetlink_dump_labels(skb, ct) < 0 ||
 531             ctnetlink_dump_ct_seq_adj(skb, ct) < 0 ||
 532             ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 533                 return -1;
 534
 535         return 0;
 536 }
 537
 538 static int ctnetlink_dump_info(struct sk_buff *skb, struct nf_conn *ct)
 539 {
 540         if (ctnetlink_dump_status(skb, ct) < 0 ||
 541             ctnetlink_dump_mark(skb, ct, true) < 0 ||
 542             ctnetlink_dump_secctx(skb, ct) < 0 ||
 543             ctnetlink_dump_id(skb, ct) < 0 ||
 544             ctnetlink_dump_use(skb, ct) < 0 ||
 545             ctnetlink_dump_master(skb, ct) < 0)
 546                 return -1;
 547
 548         if (!test_bit(IPS_OFFLOAD_BIT, &ct->status) &&
 549             (ctnetlink_dump_timeout(skb, ct) < 0 ||
 550              ctnetlink_dump_protoinfo(skb, ct) < 0))
 551                 return -1;
 552
 553         return 0;
 554 }
 555
 556 static int
 557 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 558                     struct nf_conn *ct, bool extinfo, unsigned int flags)
 559 {
 560         const struct nf_conntrack_zone *zone;
 561         struct nlmsghdr *nlh;
 562         struct nlattr *nest_parms;
 563         unsigned int event;
 564
 565         if (portid)
 566                 flags |= NLM_F_MULTI;
 567         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_NEW);
 568         nlh = nfnl_msg_put(skb, portid, seq, event, flags, nf_ct_l3num(ct),
 569                            NFNETLINK_V0, 0);
 570         if (!nlh)
 571                 goto nlmsg_failure;
 572
 573         zone = nf_ct_zone(ct);
 574
 575         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 576         if (!nest_parms)
 577                 goto nla_put_failure;
 578         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 579                 goto nla_put_failure;
 580         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 581                                    NF_CT_ZONE_DIR_ORIG) < 0)
 582                 goto nla_put_failure;
 583         nla_nest_end(skb, nest_parms);
 584
 585         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 586         if (!nest_parms)
 587                 goto nla_put_failure;
 588         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 589                 goto nla_put_failure;
 590         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 591                                    NF_CT_ZONE_DIR_REPL) < 0)
 592                 goto nla_put_failure;
 593         nla_nest_end(skb, nest_parms);
 594
 595         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 596                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 597                 goto nla_put_failure;
 598
 599         if (ctnetlink_dump_info(skb, ct) < 0)
 600                 goto nla_put_failure;
 601         if (extinfo && ctnetlink_dump_extinfo(skb, ct, type) < 0)
 602                 goto nla_put_failure;
 603
 604         nlmsg_end(skb, nlh);
 605         return skb->len;
 606
 607 nlmsg_failure:
 608 nla_put_failure:
 609         nlmsg_cancel(skb, nlh);
 610         return -1;
 611 }
 612
 613 static const struct nla_policy cta_ip_nla_policy[CTA_IP_MAX + 1] = {
 614         [CTA_IP_V4_SRC] = { .type = NLA_U32 },
 615         [CTA_IP_V4_DST] = { .type = NLA_U32 },
 616         [CTA_IP_V6_SRC] = { .len = sizeof(__be32) * 4 },
 617         [CTA_IP_V6_DST] = { .len = sizeof(__be32) * 4 },
 618 };
 619
 620 #if defined(CONFIG_NETFILTER_NETLINK_GLUE_CT) || defined(CONFIG_NF_CONNTRACK_EVENTS)
 621 static size_t ctnetlink_proto_size(const struct nf_conn *ct)
 622 {
 623         const struct nf_conntrack_l4proto *l4proto;
 624         size_t len, len4 = 0;
 625
 626         len = nla_policy_len(cta_ip_nla_policy, CTA_IP_MAX + 1);
 627         len *= 3u; /* ORIG, REPLY, MASTER */
 628
 629         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
 630         len += l4proto->nlattr_size;
 631         if (l4proto->nlattr_tuple_size) {
 632                 len4 = l4proto->nlattr_tuple_size();
 633                 len4 *= 3u; /* ORIG, REPLY, MASTER */
 634         }
 635
 636         return len + len4;
 637 }
 638 #endif
 639
 640 static inline size_t ctnetlink_acct_size(const struct nf_conn *ct)
 641 {
 642         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 643                 return 0;
 644         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 645                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 646                + 2 * nla_total_size_64bit(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 647                ;
 648 }
 649
 650 static inline int ctnetlink_secctx_size(const struct nf_conn *ct)
 651 {
 652 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 653         int len, ret;
 654
 655         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 656         if (ret)
 657                 return 0;
 658
 659         return nla_total_size(0) /* CTA_SECCTX */
 660                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 661 #else
 662         return 0;
 663 #endif
 664 }
 665
 666 static inline size_t ctnetlink_timestamp_size(const struct nf_conn *ct)
 667 {
 668 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 669         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 670                 return 0;
 671         return nla_total_size(0) + 2 * nla_total_size_64bit(sizeof(uint64_t));
 672 #else
 673         return 0;
 674 #endif
 675 }
 676
 677 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 678 static size_t ctnetlink_nlmsg_size(const struct nf_conn *ct)
 679 {
 680         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 681                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 682                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 683                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 684                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 685                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 686                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 687                + ctnetlink_acct_size(ct)
 688                + ctnetlink_timestamp_size(ct)
 689                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 690                + nla_total_size(0) /* CTA_PROTOINFO */
 691                + nla_total_size(0) /* CTA_HELP */
 692                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 693                + ctnetlink_secctx_size(ct)
 694 #if IS_ENABLED(CONFIG_NF_NAT)
 695                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 696                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 697 #endif
 698 #ifdef CONFIG_NF_CONNTRACK_MARK
 699                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 700 #endif
 701 #ifdef CONFIG_NF_CONNTRACK_ZONES
 702                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
 703 #endif
 704                + ctnetlink_proto_size(ct)
 705                + ctnetlink_label_size(ct)
 706                ;
 707 }
 708
 709 static int
 710 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 711 {
 712         const struct nf_conntrack_zone *zone;
 713         struct net *net;
 714         struct nlmsghdr *nlh;
 715         struct nlattr *nest_parms;
 716         struct nf_conn *ct = item->ct;
 717         struct sk_buff *skb;
 718         unsigned int type;
 719         unsigned int flags = 0, group;
 720         int err;
 721
 722         if (events & (1 << IPCT_DESTROY)) {
 723                 type = IPCTNL_MSG_CT_DELETE;
 724                 group = NFNLGRP_CONNTRACK_DESTROY;
 725         } else if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 726                 type = IPCTNL_MSG_CT_NEW;
 727                 flags = NLM_F_CREATE|NLM_F_EXCL;
 728                 group = NFNLGRP_CONNTRACK_NEW;
 729         } else if (events) {
 730                 type = IPCTNL_MSG_CT_NEW;
 731                 group = NFNLGRP_CONNTRACK_UPDATE;
 732         } else
 733                 return 0;
 734
 735         net = nf_ct_net(ct);
 736         if (!item->report && !nfnetlink_has_listeners(net, group))
 737                 return 0;
 738
 739         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 740         if (skb == NULL)
 741                 goto errout;
 742
 743         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, type);
 744         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags, nf_ct_l3num(ct),
 745                            NFNETLINK_V0, 0);
 746         if (!nlh)
 747                 goto nlmsg_failure;
 748
 749         zone = nf_ct_zone(ct);
 750
 751         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
 752         if (!nest_parms)
 753                 goto nla_put_failure;
 754         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 755                 goto nla_put_failure;
 756         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 757                                    NF_CT_ZONE_DIR_ORIG) < 0)
 758                 goto nla_put_failure;
 759         nla_nest_end(skb, nest_parms);
 760
 761         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
 762         if (!nest_parms)
 763                 goto nla_put_failure;
 764         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 765                 goto nla_put_failure;
 766         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
 767                                    NF_CT_ZONE_DIR_REPL) < 0)
 768                 goto nla_put_failure;
 769         nla_nest_end(skb, nest_parms);
 770
 771         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
 772                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
 773                 goto nla_put_failure;
 774
 775         if (ctnetlink_dump_id(skb, ct) < 0)
 776                 goto nla_put_failure;
 777
 778         if (ctnetlink_dump_status(skb, ct) < 0)
 779                 goto nla_put_failure;
 780
 781         if (events & (1 << IPCT_DESTROY)) {
 782                 if (ctnetlink_dump_acct(skb, ct, type) < 0 ||
 783                     ctnetlink_dump_timestamp(skb, ct) < 0)
 784                         goto nla_put_failure;
 785         } else {
 786                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 787                         goto nla_put_failure;
 788
 789                 if (events & (1 << IPCT_PROTOINFO)
 790                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 791                         goto nla_put_failure;
 792
 793                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 794                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 795                         goto nla_put_failure;
 796
 797 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 798                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 799                     && ctnetlink_dump_secctx(skb, ct) < 0)
 800                         goto nla_put_failure;
 801 #endif
 802                 if (events & (1 << IPCT_LABEL) &&
 803                      ctnetlink_dump_labels(skb, ct) < 0)
 804                         goto nla_put_failure;
 805
 806                 if (events & (1 << IPCT_RELATED) &&
 807                     ctnetlink_dump_master(skb, ct) < 0)
 808                         goto nla_put_failure;
 809
 810                 if (events & (1 << IPCT_SEQADJ) &&
 811                     ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
 812                         goto nla_put_failure;
 813
 814                 if (events & (1 << IPCT_SYNPROXY) &&
 815                     ctnetlink_dump_ct_synproxy(skb, ct) < 0)
 816                         goto nla_put_failure;
 817         }
 818
 819 #ifdef CONFIG_NF_CONNTRACK_MARK
 820         if (ctnetlink_dump_mark(skb, ct, events & (1 << IPCT_MARK)))
 821                 goto nla_put_failure;
 822 #endif
 823         nlmsg_end(skb, nlh);
 824         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 825                              GFP_ATOMIC);
 826         if (err == -ENOBUFS || err == -EAGAIN)
 827                 return -ENOBUFS;
 828
 829         return 0;
 830
 831 nla_put_failure:
 832         nlmsg_cancel(skb, nlh);
 833 nlmsg_failure:
 834         kfree_skb(skb);
 835 errout:
 836         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 837                 return -ENOBUFS;
 838
 839         return 0;
 840 }
 841 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 842
 843 static int ctnetlink_done(struct netlink_callback *cb)
 844 {
 845         if (cb->args[1])
 846                 nf_ct_put((struct nf_conn *)cb->args[1]);
 847         kfree(cb->data);
 848         return 0;
 849 }
 850
 851 struct ctnetlink_filter {
 852         u8 family;
 853
 854         u_int32_t orig_flags;
 855         u_int32_t reply_flags;
 856
 857         struct nf_conntrack_tuple orig;
 858         struct nf_conntrack_tuple reply;
 859         struct nf_conntrack_zone zone;
 860
 861         struct {
 862                 u_int32_t val;
 863                 u_int32_t mask;
 864         } mark;
 865 };
 866
 867 static const struct nla_policy cta_filter_nla_policy[CTA_FILTER_MAX + 1] = {
 868         [CTA_FILTER_ORIG_FLAGS]         = { .type = NLA_U32 },
 869         [CTA_FILTER_REPLY_FLAGS]        = { .type = NLA_U32 },
 870 };
 871
 872 static int ctnetlink_parse_filter(const struct nlattr *attr,
 873                                   struct ctnetlink_filter *filter)
 874 {
 875         struct nlattr *tb[CTA_FILTER_MAX + 1];
 876         int ret = 0;
 877
 878         ret = nla_parse_nested(tb, CTA_FILTER_MAX, attr, cta_filter_nla_policy,
 879                                NULL);
 880         if (ret)
 881                 return ret;
 882
 883         if (tb[CTA_FILTER_ORIG_FLAGS]) {
 884                 filter->orig_flags = nla_get_u32(tb[CTA_FILTER_ORIG_FLAGS]);
 885                 if (filter->orig_flags & ~CTA_FILTER_F_ALL)
 886                         return -EOPNOTSUPP;
 887         }
 888
 889         if (tb[CTA_FILTER_REPLY_FLAGS]) {
 890                 filter->reply_flags = nla_get_u32(tb[CTA_FILTER_REPLY_FLAGS]);
 891                 if (filter->reply_flags & ~CTA_FILTER_F_ALL)
 892                         return -EOPNOTSUPP;
 893         }
 894
 895         return 0;
 896 }
 897
 898 static int ctnetlink_parse_zone(const struct nlattr *attr,
 899                                 struct nf_conntrack_zone *zone);
 900 static int ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
 901                                          struct nf_conntrack_tuple *tuple,
 902                                          u32 type, u_int8_t l3num,
 903                                          struct nf_conntrack_zone *zone,
 904                                          u_int32_t flags);
 905
 906 static struct ctnetlink_filter *
 907 ctnetlink_alloc_filter(const struct nlattr * const cda[], u8 family)
 908 {
 909         struct ctnetlink_filter *filter;
 910         int err;
 911
 912 #ifndef CONFIG_NF_CONNTRACK_MARK
 913         if (cda[CTA_MARK] || cda[CTA_MARK_MASK])
 914                 return ERR_PTR(-EOPNOTSUPP);
 915 #endif
 916
 917         filter = kzalloc(sizeof(*filter), GFP_KERNEL);
 918         if (filter == NULL)
 919                 return ERR_PTR(-ENOMEM);
 920
 921         filter->family = family;
 922
 923 #ifdef CONFIG_NF_CONNTRACK_MARK
 924         if (cda[CTA_MARK]) {
 925                 filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
 926                 if (cda[CTA_MARK_MASK])
 927                         filter->mark.mask = ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
 928                 else
 929                         filter->mark.mask = 0xffffffff;
 930         } else if (cda[CTA_MARK_MASK]) {
 931                 err = -EINVAL;
 932                 goto err_filter;
 933         }
 934 #endif
 935         if (!cda[CTA_FILTER])
 936                 return filter;
 937
 938         err = ctnetlink_parse_zone(cda[CTA_ZONE], &filter->zone);
 939         if (err < 0)
 940                 goto err_filter;
 941
 942         err = ctnetlink_parse_filter(cda[CTA_FILTER], filter);
 943         if (err < 0)
 944                 goto err_filter;
 945
 946         if (filter->orig_flags) {
 947                 if (!cda[CTA_TUPLE_ORIG]) {
 948                         err = -EINVAL;
 949                         goto err_filter;
 950                 }
 951
 952                 err = ctnetlink_parse_tuple_filter(cda, &filter->orig,
 953                                                    CTA_TUPLE_ORIG,
 954                                                    filter->family,
 955                                                    &filter->zone,
 956                                                    filter->orig_flags);
 957                 if (err < 0)
 958                         goto err_filter;
 959         }
 960
 961         if (filter->reply_flags) {
 962                 if (!cda[CTA_TUPLE_REPLY]) {
 963                         err = -EINVAL;
 964                         goto err_filter;
 965                 }
 966
 967                 err = ctnetlink_parse_tuple_filter(cda, &filter->reply,
 968                                                    CTA_TUPLE_REPLY,
 969                                                    filter->family,
 970                                                    &filter->zone,
 971                                                    filter->reply_flags);
 972                 if (err < 0)
 973                         goto err_filter;
 974         }
 975
 976         return filter;
 977
 978 err_filter:
 979         kfree(filter);
 980
 981         return ERR_PTR(err);
 982 }
 983
 984 static bool ctnetlink_needs_filter(u8 family, const struct nlattr * const *cda)
 985 {
 986         return family || cda[CTA_MARK] || cda[CTA_FILTER];
 987 }
 988
 989 static int ctnetlink_start(struct netlink_callback *cb)
 990 {
 991         const struct nlattr * const *cda = cb->data;
 992         struct ctnetlink_filter *filter = NULL;
 993         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 994         u8 family = nfmsg->nfgen_family;
 995
 996         if (ctnetlink_needs_filter(family, cda)) {
 997                 filter = ctnetlink_alloc_filter(cda, family);
 998                 if (IS_ERR(filter))
 999                         return PTR_ERR(filter);
1000         }
1001
1002         cb->data = filter;
1003         return 0;
1004 }
1005
1006 static int ctnetlink_filter_match_tuple(struct nf_conntrack_tuple *filter_tuple,
1007                                         struct nf_conntrack_tuple *ct_tuple,
1008                                         u_int32_t flags, int family)
1009 {
1010         switch (family) {
1011         case NFPROTO_IPV4:
1012                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1013                     filter_tuple->src.u3.ip != ct_tuple->src.u3.ip)
1014                         return  0;
1015
1016                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1017                     filter_tuple->dst.u3.ip != ct_tuple->dst.u3.ip)
1018                         return  0;
1019                 break;
1020         case NFPROTO_IPV6:
1021                 if ((flags & CTA_FILTER_FLAG(CTA_IP_SRC)) &&
1022                     !ipv6_addr_cmp(&filter_tuple->src.u3.in6,
1023                                    &ct_tuple->src.u3.in6))
1024                         return 0;
1025
1026                 if ((flags & CTA_FILTER_FLAG(CTA_IP_DST)) &&
1027                     !ipv6_addr_cmp(&filter_tuple->dst.u3.in6,
1028                                    &ct_tuple->dst.u3.in6))
1029                         return 0;
1030                 break;
1031         }
1032
1033         if ((flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) &&
1034             filter_tuple->dst.protonum != ct_tuple->dst.protonum)
1035                 return 0;
1036
1037         switch (ct_tuple->dst.protonum) {
1038         case IPPROTO_TCP:
1039         case IPPROTO_UDP:
1040                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) &&
1041                     filter_tuple->src.u.tcp.port != ct_tuple->src.u.tcp.port)
1042                         return 0;
1043
1044                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) &&
1045                     filter_tuple->dst.u.tcp.port != ct_tuple->dst.u.tcp.port)
1046                         return 0;
1047                 break;
1048         case IPPROTO_ICMP:
1049                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_TYPE)) &&
1050                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1051                         return 0;
1052                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_CODE)) &&
1053                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1054                         return 0;
1055                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMP_ID)) &&
1056                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1057                         return 0;
1058                 break;
1059         case IPPROTO_ICMPV6:
1060                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_TYPE)) &&
1061                     filter_tuple->dst.u.icmp.type != ct_tuple->dst.u.icmp.type)
1062                         return 0;
1063                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_CODE)) &&
1064                     filter_tuple->dst.u.icmp.code != ct_tuple->dst.u.icmp.code)
1065                         return 0;
1066                 if ((flags & CTA_FILTER_FLAG(CTA_PROTO_ICMPV6_ID)) &&
1067                     filter_tuple->src.u.icmp.id != ct_tuple->src.u.icmp.id)
1068                         return 0;
1069                 break;
1070         }
1071
1072         return 1;
1073 }
1074
1075 static int ctnetlink_filter_match(struct nf_conn *ct, void *data)
1076 {
1077         struct ctnetlink_filter *filter = data;
1078         struct nf_conntrack_tuple *tuple;
1079
1080         if (filter == NULL)
1081                 goto out;
1082
1083         /* Match entries of a given L3 protocol number.
1084          * If it is not specified, ie. l3proto == 0,
1085          * then match everything.
1086          */
1087         if (filter->family && nf_ct_l3num(ct) != filter->family)
1088                 goto ignore_entry;
1089
1090         if (filter->orig_flags) {
1091                 tuple = nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL);
1092                 if (!ctnetlink_filter_match_tuple(&filter->orig, tuple,
1093                                                   filter->orig_flags,
1094                                                   filter->family))
1095                         goto ignore_entry;
1096         }
1097
1098         if (filter->reply_flags) {
1099                 tuple = nf_ct_tuple(ct, IP_CT_DIR_REPLY);
1100                 if (!ctnetlink_filter_match_tuple(&filter->reply, tuple,
1101                                                   filter->reply_flags,
1102                                                   filter->family))
1103                         goto ignore_entry;
1104         }
1105
1106 #ifdef CONFIG_NF_CONNTRACK_MARK
1107         if ((READ_ONCE(ct->mark) & filter->mark.mask) != filter->mark.val)
1108                 goto ignore_entry;
1109 #endif
1110
1111 out:
1112         return 1;
1113
1114 ignore_entry:
1115         return 0;
1116 }
1117
1118 static int
1119 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
1120 {
1121         unsigned int flags = cb->data ? NLM_F_DUMP_FILTERED : 0;
1122         struct net *net = sock_net(skb->sk);
1123         struct nf_conn *ct, *last;
1124         struct nf_conntrack_tuple_hash *h;
1125         struct hlist_nulls_node *n;
1126         struct nf_conn *nf_ct_evict[8];
1127         int res, i;
1128         spinlock_t *lockp;
1129
1130         last = (struct nf_conn *)cb->args[1];
1131         i = 0;
1132
1133         local_bh_disable();
1134         for (; cb->args[0] < nf_conntrack_htable_size; cb->args[0]++) {
1135 restart:
1136                 while (i) {
1137                         i--;
1138                         if (nf_ct_should_gc(nf_ct_evict[i]))
1139                                 nf_ct_kill(nf_ct_evict[i]);
1140                         nf_ct_put(nf_ct_evict[i]);
1141                 }
1142
1143                 lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
1144                 nf_conntrack_lock(lockp);
1145                 if (cb->args[0] >= nf_conntrack_htable_size) {
1146                         spin_unlock(lockp);
1147                         goto out;
1148                 }
1149                 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[cb->args[0]],
1150                                            hnnode) {
1151                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1152                                 continue;
1153                         ct = nf_ct_tuplehash_to_ctrack(h);
1154                         if (nf_ct_is_expired(ct)) {
1155                                 if (i < ARRAY_SIZE(nf_ct_evict) &&
1156                                     atomic_inc_not_zero(&ct->ct_general.use))
1157                                         nf_ct_evict[i++] = ct;
1158                                 continue;
1159                         }
1160
1161                         if (!net_eq(net, nf_ct_net(ct)))
1162                                 continue;
1163
1164                         if (cb->args[1]) {
1165                                 if (ct != last)
1166                                         continue;
1167                                 cb->args[1] = 0;
1168                         }
1169                         if (!ctnetlink_filter_match(ct, cb->data))
1170                                 continue;
1171
1172                         res =
1173                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1174                                             cb->nlh->nlmsg_seq,
1175                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1176                                             ct, true, flags);
1177                         if (res < 0) {
1178                                 nf_conntrack_get(&ct->ct_general);
1179                                 cb->args[1] = (unsigned long)ct;
1180                                 spin_unlock(lockp);
1181                                 goto out;
1182                         }
1183                 }
1184                 spin_unlock(lockp);
1185                 if (cb->args[1]) {
1186                         cb->args[1] = 0;
1187                         goto restart;
1188                 }
1189         }
1190 out:
1191         local_bh_enable();
1192         if (last) {
1193                 /* nf ct hash resize happened, now clear the leftover. */
1194                 if ((struct nf_conn *)cb->args[1] == last)
1195                         cb->args[1] = 0;
1196
1197                 nf_ct_put(last);
1198         }
1199
1200         while (i) {
1201                 i--;
1202                 if (nf_ct_should_gc(nf_ct_evict[i]))
1203                         nf_ct_kill(nf_ct_evict[i]);
1204                 nf_ct_put(nf_ct_evict[i]);
1205         }
1206
1207         return skb->len;
1208 }
1209
1210 static int ipv4_nlattr_to_tuple(struct nlattr *tb[],
1211                                 struct nf_conntrack_tuple *t,
1212                                 u_int32_t flags)
1213 {
1214         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1215                 if (!tb[CTA_IP_V4_SRC])
1216                         return -EINVAL;
1217
1218                 t->src.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_SRC]);
1219         }
1220
1221         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1222                 if (!tb[CTA_IP_V4_DST])
1223                         return -EINVAL;
1224
1225                 t->dst.u3.ip = nla_get_in_addr(tb[CTA_IP_V4_DST]);
1226         }
1227
1228         return 0;
1229 }
1230
1231 static int ipv6_nlattr_to_tuple(struct nlattr *tb[],
1232                                 struct nf_conntrack_tuple *t,
1233                                 u_int32_t flags)
1234 {
1235         if (flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1236                 if (!tb[CTA_IP_V6_SRC])
1237                         return -EINVAL;
1238
1239                 t->src.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_SRC]);
1240         }
1241
1242         if (flags & CTA_FILTER_FLAG(CTA_IP_DST)) {
1243                 if (!tb[CTA_IP_V6_DST])
1244                         return -EINVAL;
1245
1246                 t->dst.u3.in6 = nla_get_in6_addr(tb[CTA_IP_V6_DST]);
1247         }
1248
1249         return 0;
1250 }
1251
1252 static int ctnetlink_parse_tuple_ip(struct nlattr *attr,
1253                                     struct nf_conntrack_tuple *tuple,
1254                                     u_int32_t flags)
1255 {
1256         struct nlattr *tb[CTA_IP_MAX+1];
1257         int ret = 0;
1258
1259         ret = nla_parse_nested_deprecated(tb, CTA_IP_MAX, attr, NULL, NULL);
1260         if (ret < 0)
1261                 return ret;
1262
1263         ret = nla_validate_nested_deprecated(attr, CTA_IP_MAX,
1264                                              cta_ip_nla_policy, NULL);
1265         if (ret)
1266                 return ret;
1267
1268         switch (tuple->src.l3num) {
1269         case NFPROTO_IPV4:
1270                 ret = ipv4_nlattr_to_tuple(tb, tuple, flags);
1271                 break;
1272         case NFPROTO_IPV6:
1273                 ret = ipv6_nlattr_to_tuple(tb, tuple, flags);
1274                 break;
1275         }
1276
1277         return ret;
1278 }
1279
1280 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
1281         [CTA_PROTO_NUM] = { .type = NLA_U8 },
1282 };
1283
1284 static int ctnetlink_parse_tuple_proto(struct nlattr *attr,
1285                                        struct nf_conntrack_tuple *tuple,
1286                                        u_int32_t flags)
1287 {
1288         const struct nf_conntrack_l4proto *l4proto;
1289         struct nlattr *tb[CTA_PROTO_MAX+1];
1290         int ret = 0;
1291
1292         ret = nla_parse_nested_deprecated(tb, CTA_PROTO_MAX, attr,
1293                                           proto_nla_policy, NULL);
1294         if (ret < 0)
1295                 return ret;
1296
1297         if (!(flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)))
1298                 return 0;
1299
1300         if (!tb[CTA_PROTO_NUM])
1301                 return -EINVAL;
1302
1303         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
1304
1305         rcu_read_lock();
1306         l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
1307
1308         if (likely(l4proto->nlattr_to_tuple)) {
1309                 ret = nla_validate_nested_deprecated(attr, CTA_PROTO_MAX,
1310                                                      l4proto->nla_policy,
1311                                                      NULL);
1312                 if (ret == 0)
1313                         ret = l4proto->nlattr_to_tuple(tb, tuple, flags);
1314         }
1315
1316         rcu_read_unlock();
1317
1318         return ret;
1319 }
1320
1321 static int
1322 ctnetlink_parse_zone(const struct nlattr *attr,
1323                      struct nf_conntrack_zone *zone)
1324 {
1325         nf_ct_zone_init(zone, NF_CT_DEFAULT_ZONE_ID,
1326                         NF_CT_DEFAULT_ZONE_DIR, 0);
1327 #ifdef CONFIG_NF_CONNTRACK_ZONES
1328         if (attr)
1329                 zone->id = ntohs(nla_get_be16(attr));
1330 #else
1331         if (attr)
1332                 return -EOPNOTSUPP;
1333 #endif
1334         return 0;
1335 }
1336
1337 static int
1338 ctnetlink_parse_tuple_zone(struct nlattr *attr, enum ctattr_type type,
1339                            struct nf_conntrack_zone *zone)
1340 {
1341         int ret;
1342
1343         if (zone->id != NF_CT_DEFAULT_ZONE_ID)
1344                 return -EINVAL;
1345
1346         ret = ctnetlink_parse_zone(attr, zone);
1347         if (ret < 0)
1348                 return ret;
1349
1350         if (type == CTA_TUPLE_REPLY)
1351                 zone->dir = NF_CT_ZONE_DIR_REPL;
1352         else
1353                 zone->dir = NF_CT_ZONE_DIR_ORIG;
1354
1355         return 0;
1356 }
1357
1358 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
1359         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
1360         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
1361         [CTA_TUPLE_ZONE]        = { .type = NLA_U16 },
1362 };
1363
1364 #define CTA_FILTER_F_ALL_CTA_PROTO \
1365   (CTA_FILTER_F_CTA_PROTO_SRC_PORT | \
1366    CTA_FILTER_F_CTA_PROTO_DST_PORT | \
1367    CTA_FILTER_F_CTA_PROTO_ICMP_TYPE | \
1368    CTA_FILTER_F_CTA_PROTO_ICMP_CODE | \
1369    CTA_FILTER_F_CTA_PROTO_ICMP_ID | \
1370    CTA_FILTER_F_CTA_PROTO_ICMPV6_TYPE | \
1371    CTA_FILTER_F_CTA_PROTO_ICMPV6_CODE | \
1372    CTA_FILTER_F_CTA_PROTO_ICMPV6_ID)
1373
1374 static int
1375 ctnetlink_parse_tuple_filter(const struct nlattr * const cda[],
1376                               struct nf_conntrack_tuple *tuple, u32 type,
1377                               u_int8_t l3num, struct nf_conntrack_zone *zone,
1378                               u_int32_t flags)
1379 {
1380         struct nlattr *tb[CTA_TUPLE_MAX+1];
1381         int err;
1382
1383         memset(tuple, 0, sizeof(*tuple));
1384
1385         err = nla_parse_nested_deprecated(tb, CTA_TUPLE_MAX, cda[type],
1386                                           tuple_nla_policy, NULL);
1387         if (err < 0)
1388                 return err;
1389
1390         if (l3num != NFPROTO_IPV4 && l3num != NFPROTO_IPV6)
1391                 return -EOPNOTSUPP;
1392         tuple->src.l3num = l3num;
1393
1394         if (flags & CTA_FILTER_FLAG(CTA_IP_DST) ||
1395             flags & CTA_FILTER_FLAG(CTA_IP_SRC)) {
1396                 if (!tb[CTA_TUPLE_IP])
1397                         return -EINVAL;
1398
1399                 err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple, flags);
1400                 if (err < 0)
1401                         return err;
1402         }
1403
1404         if (flags & CTA_FILTER_FLAG(CTA_PROTO_NUM)) {
1405                 if (!tb[CTA_TUPLE_PROTO])
1406                         return -EINVAL;
1407
1408                 err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple, flags);
1409                 if (err < 0)
1410                         return err;
1411         } else if (flags & CTA_FILTER_FLAG(ALL_CTA_PROTO)) {
1412                 /* Can't manage proto flags without a protonum  */
1413                 return -EINVAL;
1414         }
1415
1416         if ((flags & CTA_FILTER_FLAG(CTA_TUPLE_ZONE)) && tb[CTA_TUPLE_ZONE]) {
1417                 if (!zone)
1418                         return -EINVAL;
1419
1420                 err = ctnetlink_parse_tuple_zone(tb[CTA_TUPLE_ZONE],
1421                                                  type, zone);
1422                 if (err < 0)
1423                         return err;
1424         }
1425
1426         /* orig and expect tuples get DIR_ORIGINAL */
1427         if (type == CTA_TUPLE_REPLY)
1428                 tuple->dst.dir = IP_CT_DIR_REPLY;
1429         else
1430                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
1431
1432         return 0;
1433 }
1434
1435 static int
1436 ctnetlink_parse_tuple(const struct nlattr * const cda[],
1437                       struct nf_conntrack_tuple *tuple, u32 type,
1438                       u_int8_t l3num, struct nf_conntrack_zone *zone)
1439 {
1440         return ctnetlink_parse_tuple_filter(cda, tuple, type, l3num, zone,
1441                                             CTA_FILTER_FLAG(ALL));
1442 }
1443
1444 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
1445         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING,
1446                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
1447 };
1448
1449 static int ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
1450                                 struct nlattr **helpinfo)
1451 {
1452         int err;
1453         struct nlattr *tb[CTA_HELP_MAX+1];
1454
1455         err = nla_parse_nested_deprecated(tb, CTA_HELP_MAX, attr,
1456                                           help_nla_policy, NULL);
1457         if (err < 0)
1458                 return err;
1459
1460         if (!tb[CTA_HELP_NAME])
1461                 return -EINVAL;
1462
1463         *helper_name = nla_data(tb[CTA_HELP_NAME]);
1464
1465         if (tb[CTA_HELP_INFO])
1466                 *helpinfo = tb[CTA_HELP_INFO];
1467
1468         return 0;
1469 }
1470
1471 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
1472         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
1473         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
1474         [CTA_STATUS]            = { .type = NLA_U32 },
1475         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
1476         [CTA_HELP]              = { .type = NLA_NESTED },
1477         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
1478         [CTA_TIMEOUT]           = { .type = NLA_U32 },
1479         [CTA_MARK]              = { .type = NLA_U32 },
1480         [CTA_ID]                = { .type = NLA_U32 },
1481         [CTA_NAT_DST]           = { .type = NLA_NESTED },
1482         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
1483         [CTA_NAT_SEQ_ADJ_ORIG]  = { .type = NLA_NESTED },
1484         [CTA_NAT_SEQ_ADJ_REPLY] = { .type = NLA_NESTED },
1485         [CTA_ZONE]              = { .type = NLA_U16 },
1486         [CTA_MARK_MASK]         = { .type = NLA_U32 },
1487         [CTA_LABELS]            = { .type = NLA_BINARY,
1488                                     .len = NF_CT_LABELS_MAX_SIZE },
1489         [CTA_LABELS_MASK]       = { .type = NLA_BINARY,
1490                                     .len = NF_CT_LABELS_MAX_SIZE },
1491         [CTA_FILTER]            = { .type = NLA_NESTED },
1492 };
1493
1494 static int ctnetlink_flush_iterate(struct nf_conn *ct, void *data)
1495 {
1496         return ctnetlink_filter_match(ct, data);
1497 }
1498
1499 static int ctnetlink_flush_conntrack(struct net *net,
1500                                      const struct nlattr * const cda[],
1501                                      u32 portid, int report, u8 family)
1502 {
1503         struct ctnetlink_filter *filter = NULL;
1504
1505         if (ctnetlink_needs_filter(family, cda)) {
1506                 if (cda[CTA_FILTER])
1507                         return -EOPNOTSUPP;
1508
1509                 filter = ctnetlink_alloc_filter(cda, family);
1510                 if (IS_ERR(filter))
1511                         return PTR_ERR(filter);
1512         }
1513
1514         nf_ct_iterate_cleanup_net(net, ctnetlink_flush_iterate, filter,
1515                                   portid, report);
1516         kfree(filter);
1517
1518         return 0;
1519 }
1520
1521 static int ctnetlink_del_conntrack(struct net *net, struct sock *ctnl,
1522                                    struct sk_buff *skb,
1523                                    const struct nlmsghdr *nlh,
1524                                    const struct nlattr * const cda[],
1525                                    struct netlink_ext_ack *extack)
1526 {
1527         struct nf_conntrack_tuple_hash *h;
1528         struct nf_conntrack_tuple tuple;
1529         struct nf_conn *ct;
1530         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1531         struct nf_conntrack_zone zone;
1532         int err;
1533
1534         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1535         if (err < 0)
1536                 return err;
1537
1538         if (cda[CTA_TUPLE_ORIG])
1539                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1540                                             nfmsg->nfgen_family, &zone);
1541         else if (cda[CTA_TUPLE_REPLY])
1542                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1543                                             nfmsg->nfgen_family, &zone);
1544         else {
1545                 u_int8_t u3 = nfmsg->version ? nfmsg->nfgen_family : AF_UNSPEC;
1546
1547                 return ctnetlink_flush_conntrack(net, cda,
1548                                                  NETLINK_CB(skb).portid,
1549                                                  nlmsg_report(nlh), u3);
1550         }
1551
1552         if (err < 0)
1553                 return err;
1554
1555         h = nf_conntrack_find_get(net, &zone, &tuple);
1556         if (!h)
1557                 return -ENOENT;
1558
1559         ct = nf_ct_tuplehash_to_ctrack(h);
1560
1561         if (cda[CTA_ID]) {
1562                 __be32 id = nla_get_be32(cda[CTA_ID]);
1563
1564                 if (id != (__force __be32)nf_ct_get_id(ct)) {
1565                         nf_ct_put(ct);
1566                         return -ENOENT;
1567                 }
1568         }
1569
1570         nf_ct_delete(ct, NETLINK_CB(skb).portid, nlmsg_report(nlh));
1571         nf_ct_put(ct);
1572
1573         return 0;
1574 }
1575
1576 static int ctnetlink_get_conntrack(struct net *net, struct sock *ctnl,
1577                                    struct sk_buff *skb,
1578                                    const struct nlmsghdr *nlh,
1579                                    const struct nlattr * const cda[],
1580                                    struct netlink_ext_ack *extack)
1581 {
1582         struct nf_conntrack_tuple_hash *h;
1583         struct nf_conntrack_tuple tuple;
1584         struct nf_conn *ct;
1585         struct sk_buff *skb2 = NULL;
1586         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1587         u_int8_t u3 = nfmsg->nfgen_family;
1588         struct nf_conntrack_zone zone;
1589         int err;
1590
1591         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1592                 struct netlink_dump_control c = {
1593                         .start = ctnetlink_start,
1594                         .dump = ctnetlink_dump_table,
1595                         .done = ctnetlink_done,
1596                         .data = (void *)cda,
1597                 };
1598
1599                 return netlink_dump_start(ctnl, skb, nlh, &c);
1600         }
1601
1602         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1603         if (err < 0)
1604                 return err;
1605
1606         if (cda[CTA_TUPLE_ORIG])
1607                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG,
1608                                             u3, &zone);
1609         else if (cda[CTA_TUPLE_REPLY])
1610                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY,
1611                                             u3, &zone);
1612         else
1613                 return -EINVAL;
1614
1615         if (err < 0)
1616                 return err;
1617
1618         h = nf_conntrack_find_get(net, &zone, &tuple);
1619         if (!h)
1620                 return -ENOENT;
1621
1622         ct = nf_ct_tuplehash_to_ctrack(h);
1623
1624         err = -ENOMEM;
1625         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1626         if (skb2 == NULL) {
1627                 nf_ct_put(ct);
1628                 return -ENOMEM;
1629         }
1630
1631         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1632                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct, true, 0);
1633         nf_ct_put(ct);
1634         if (err <= 0)
1635                 goto free;
1636
1637         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1638         if (err < 0)
1639                 goto out;
1640
1641         return 0;
1642
1643 free:
1644         kfree_skb(skb2);
1645 out:
1646         /* this avoids a loop in nfnetlink. */
1647         return err == -EAGAIN ? -ENOBUFS : err;
1648 }
1649
1650 static int ctnetlink_done_list(struct netlink_callback *cb)
1651 {
1652         if (cb->args[1])
1653                 nf_ct_put((struct nf_conn *)cb->args[1]);
1654         return 0;
1655 }
1656
1657 static int
1658 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb, bool dying)
1659 {
1660         struct nf_conn *ct, *last;
1661         struct nf_conntrack_tuple_hash *h;
1662         struct hlist_nulls_node *n;
1663         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1664         u_int8_t l3proto = nfmsg->nfgen_family;
1665         int res;
1666         int cpu;
1667         struct hlist_nulls_head *list;
1668         struct net *net = sock_net(skb->sk);
1669
1670         if (cb->args[2])
1671                 return 0;
1672
1673         last = (struct nf_conn *)cb->args[1];
1674
1675         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1676                 struct ct_pcpu *pcpu;
1677
1678                 if (!cpu_possible(cpu))
1679                         continue;
1680
1681                 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1682                 spin_lock_bh(&pcpu->lock);
1683                 list = dying ? &pcpu->dying : &pcpu->unconfirmed;
1684 restart:
1685                 hlist_nulls_for_each_entry(h, n, list, hnnode) {
1686                         ct = nf_ct_tuplehash_to_ctrack(h);
1687                         if (l3proto && nf_ct_l3num(ct) != l3proto)
1688                                 continue;
1689                         if (cb->args[1]) {
1690                                 if (ct != last)
1691                                         continue;
1692                                 cb->args[1] = 0;
1693                         }
1694
1695                         /* We can't dump extension info for the unconfirmed
1696                          * list because unconfirmed conntracks can have
1697                          * ct->ext reallocated (and thus freed).
1698                          *
1699                          * In the dying list case ct->ext can't be free'd
1700                          * until after we drop pcpu->lock.
1701                          */
1702                         res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1703                                                   cb->nlh->nlmsg_seq,
1704                                                   NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1705                                                   ct, dying ? true : false, 0);
1706                         if (res < 0) {
1707                                 if (!atomic_inc_not_zero(&ct->ct_general.use))
1708                                         continue;
1709                                 cb->args[0] = cpu;
1710                                 cb->args[1] = (unsigned long)ct;
1711                                 spin_unlock_bh(&pcpu->lock);
1712                                 goto out;
1713                         }
1714                 }
1715                 if (cb->args[1]) {
1716                         cb->args[1] = 0;
1717                         goto restart;
1718                 }
1719                 spin_unlock_bh(&pcpu->lock);
1720         }
1721         cb->args[2] = 1;
1722 out:
1723         if (last)
1724                 nf_ct_put(last);
1725
1726         return skb->len;
1727 }
1728
1729 static int
1730 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1731 {
1732         return ctnetlink_dump_list(skb, cb, true);
1733 }
1734
1735 static int ctnetlink_get_ct_dying(struct net *net, struct sock *ctnl,
1736                                   struct sk_buff *skb,
1737                                   const struct nlmsghdr *nlh,
1738                                   const struct nlattr * const cda[],
1739                                   struct netlink_ext_ack *extack)
1740 {
1741         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1742                 struct netlink_dump_control c = {
1743                         .dump = ctnetlink_dump_dying,
1744                         .done = ctnetlink_done_list,
1745                 };
1746                 return netlink_dump_start(ctnl, skb, nlh, &c);
1747         }
1748
1749         return -EOPNOTSUPP;
1750 }
1751
1752 static int
1753 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1754 {
1755         return ctnetlink_dump_list(skb, cb, false);
1756 }
1757
1758 static int ctnetlink_get_ct_unconfirmed(struct net *net, struct sock *ctnl,
1759                                         struct sk_buff *skb,
1760                                         const struct nlmsghdr *nlh,
1761                                         const struct nlattr * const cda[],
1762                                         struct netlink_ext_ack *extack)
1763 {
1764         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1765                 struct netlink_dump_control c = {
1766                         .dump = ctnetlink_dump_unconfirmed,
1767                         .done = ctnetlink_done_list,
1768                 };
1769                 return netlink_dump_start(ctnl, skb, nlh, &c);
1770         }
1771
1772         return -EOPNOTSUPP;
1773 }
1774
1775 #if IS_ENABLED(CONFIG_NF_NAT)
1776 static int
1777 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1778                           enum nf_nat_manip_type manip,
1779                           const struct nlattr *attr)
1780         __must_hold(RCU)
1781 {
1782         struct nf_nat_hook *nat_hook;
1783         int err;
1784
1785         nat_hook = rcu_dereference(nf_nat_hook);
1786         if (!nat_hook) {
1787 #ifdef CONFIG_MODULES
1788                 rcu_read_unlock();
1789                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1790                 if (request_module("nf-nat") < 0) {
1791                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1792                         rcu_read_lock();
1793                         return -EOPNOTSUPP;
1794                 }
1795                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1796                 rcu_read_lock();
1797                 nat_hook = rcu_dereference(nf_nat_hook);
1798                 if (nat_hook)
1799                         return -EAGAIN;
1800 #endif
1801                 return -EOPNOTSUPP;
1802         }
1803
1804         err = nat_hook->parse_nat_setup(ct, manip, attr);
1805         if (err == -EAGAIN) {
1806 #ifdef CONFIG_MODULES
1807                 rcu_read_unlock();
1808                 nfnl_unlock(NFNL_SUBSYS_CTNETLINK);
1809                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1810                         nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1811                         rcu_read_lock();
1812                         return -EOPNOTSUPP;
1813                 }
1814                 nfnl_lock(NFNL_SUBSYS_CTNETLINK);
1815                 rcu_read_lock();
1816 #else
1817                 err = -EOPNOTSUPP;
1818 #endif
1819         }
1820         return err;
1821 }
1822 #endif
1823
1824 static void
1825 __ctnetlink_change_status(struct nf_conn *ct, unsigned long on,
1826                           unsigned long off)
1827 {
1828         unsigned int bit;
1829
1830         /* Ignore these unchangable bits */
1831         on &= ~IPS_UNCHANGEABLE_MASK;
1832         off &= ~IPS_UNCHANGEABLE_MASK;
1833
1834         for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
1835                 if (on & (1 << bit))
1836                         set_bit(bit, &ct->status);
1837                 else if (off & (1 << bit))
1838                         clear_bit(bit, &ct->status);
1839         }
1840 }
1841
1842 static int
1843 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1844 {
1845         unsigned long d;
1846         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1847         d = ct->status ^ status;
1848
1849         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1850                 /* unchangeable */
1851                 return -EBUSY;
1852
1853         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1854                 /* SEEN_REPLY bit can only be set */
1855                 return -EBUSY;
1856
1857         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1858                 /* ASSURED bit can only be set */
1859                 return -EBUSY;
1860
1861         __ctnetlink_change_status(ct, status, 0);
1862         return 0;
1863 }
1864
1865 static int
1866 ctnetlink_setup_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1867 {
1868 #if IS_ENABLED(CONFIG_NF_NAT)
1869         int ret;
1870
1871         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1872                 return 0;
1873
1874         ret = ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_DST,
1875                                         cda[CTA_NAT_DST]);
1876         if (ret < 0)
1877                 return ret;
1878
1879         return ctnetlink_parse_nat_setup(ct, NF_NAT_MANIP_SRC,
1880                                          cda[CTA_NAT_SRC]);
1881 #else
1882         if (!cda[CTA_NAT_DST] && !cda[CTA_NAT_SRC])
1883                 return 0;
1884         return -EOPNOTSUPP;
1885 #endif
1886 }
1887
1888 static int ctnetlink_change_helper(struct nf_conn *ct,
1889                                    const struct nlattr * const cda[])
1890 {
1891         struct nf_conntrack_helper *helper;
1892         struct nf_conn_help *help = nfct_help(ct);
1893         char *helpname = NULL;
1894         struct nlattr *helpinfo = NULL;
1895         int err;
1896
1897         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1898         if (err < 0)
1899                 return err;
1900
1901         /* don't change helper of sibling connections */
1902         if (ct->master) {
1903                 /* If we try to change the helper to the same thing twice,
1904                  * treat the second attempt as a no-op instead of returning
1905                  * an error.
1906                  */
1907                 err = -EBUSY;
1908                 if (help) {
1909                         rcu_read_lock();
1910                         helper = rcu_dereference(help->helper);
1911                         if (helper && !strcmp(helper->name, helpname))
1912                                 err = 0;
1913                         rcu_read_unlock();
1914                 }
1915
1916                 return err;
1917         }
1918
1919         if (!strcmp(helpname, "")) {
1920                 if (help && help->helper) {
1921                         /* we had a helper before ... */
1922                         nf_ct_remove_expectations(ct);
1923                         RCU_INIT_POINTER(help->helper, NULL);
1924                 }
1925
1926                 return 0;
1927         }
1928
1929         rcu_read_lock();
1930         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1931                                             nf_ct_protonum(ct));
1932         if (helper == NULL) {
1933                 rcu_read_unlock();
1934                 return -EOPNOTSUPP;
1935         }
1936
1937         if (help) {
1938                 if (help->helper == helper) {
1939                         /* update private helper data if allowed. */
1940                         if (helper->from_nlattr)
1941                                 helper->from_nlattr(helpinfo, ct);
1942                         err = 0;
1943                 } else
1944                         err = -EBUSY;
1945         } else {
1946                 /* we cannot set a helper for an existing conntrack */
1947                 err = -EOPNOTSUPP;
1948         }
1949
1950         rcu_read_unlock();
1951         return err;
1952 }
1953
1954 static int ctnetlink_change_timeout(struct nf_conn *ct,
1955                                     const struct nlattr * const cda[])
1956 {
1957         u64 timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
1958
1959         if (timeout > INT_MAX)
1960                 timeout = INT_MAX;
1961         WRITE_ONCE(ct->timeout, nfct_time_stamp + (u32)timeout);
1962
1963         if (test_bit(IPS_DYING_BIT, &ct->status))
1964                 return -ETIME;
1965
1966         return 0;
1967 }
1968
1969 #if defined(CONFIG_NF_CONNTRACK_MARK)
1970 static void ctnetlink_change_mark(struct nf_conn *ct,
1971                                     const struct nlattr * const cda[])
1972 {
1973         u32 mark, newmark, mask = 0;
1974
1975         if (cda[CTA_MARK_MASK])
1976                 mask = ~ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1977
1978         mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1979         newmark = (READ_ONCE(ct->mark) & mask) ^ mark;
1980         if (newmark != READ_ONCE(ct->mark))
1981                 WRITE_ONCE(ct->mark, newmark);
1982 }
1983 #endif
1984
1985 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1986         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1987         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1988         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1989 };
1990
1991 static int ctnetlink_change_protoinfo(struct nf_conn *ct,
1992                                       const struct nlattr * const cda[])
1993 {
1994         const struct nlattr *attr = cda[CTA_PROTOINFO];
1995         const struct nf_conntrack_l4proto *l4proto;
1996         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1997         int err = 0;
1998
1999         err = nla_parse_nested_deprecated(tb, CTA_PROTOINFO_MAX, attr,
2000                                           protoinfo_policy, NULL);
2001         if (err < 0)
2002                 return err;
2003
2004         l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
2005         if (l4proto->from_nlattr)
2006                 err = l4proto->from_nlattr(tb, ct);
2007
2008         return err;
2009 }
2010
2011 static const struct nla_policy seqadj_policy[CTA_SEQADJ_MAX+1] = {
2012         [CTA_SEQADJ_CORRECTION_POS]     = { .type = NLA_U32 },
2013         [CTA_SEQADJ_OFFSET_BEFORE]      = { .type = NLA_U32 },
2014         [CTA_SEQADJ_OFFSET_AFTER]       = { .type = NLA_U32 },
2015 };
2016
2017 static int change_seq_adj(struct nf_ct_seqadj *seq,
2018                           const struct nlattr * const attr)
2019 {
2020         int err;
2021         struct nlattr *cda[CTA_SEQADJ_MAX+1];
2022
2023         err = nla_parse_nested_deprecated(cda, CTA_SEQADJ_MAX, attr,
2024                                           seqadj_policy, NULL);
2025         if (err < 0)
2026                 return err;
2027
2028         if (!cda[CTA_SEQADJ_CORRECTION_POS])
2029                 return -EINVAL;
2030
2031         seq->correction_pos =
2032                 ntohl(nla_get_be32(cda[CTA_SEQADJ_CORRECTION_POS]));
2033
2034         if (!cda[CTA_SEQADJ_OFFSET_BEFORE])
2035                 return -EINVAL;
2036
2037         seq->offset_before =
2038                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_BEFORE]));
2039
2040         if (!cda[CTA_SEQADJ_OFFSET_AFTER])
2041                 return -EINVAL;
2042
2043         seq->offset_after =
2044                 ntohl(nla_get_be32(cda[CTA_SEQADJ_OFFSET_AFTER]));
2045
2046         return 0;
2047 }
2048
2049 static int
2050 ctnetlink_change_seq_adj(struct nf_conn *ct,
2051                          const struct nlattr * const cda[])
2052 {
2053         struct nf_conn_seqadj *seqadj = nfct_seqadj(ct);
2054         int ret = 0;
2055
2056         if (!seqadj)
2057                 return 0;
2058
2059         spin_lock_bh(&ct->lock);
2060         if (cda[CTA_SEQ_ADJ_ORIG]) {
2061                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_ORIGINAL],
2062                                      cda[CTA_SEQ_ADJ_ORIG]);
2063                 if (ret < 0)
2064                         goto err;
2065
2066                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2067         }
2068
2069         if (cda[CTA_SEQ_ADJ_REPLY]) {
2070                 ret = change_seq_adj(&seqadj->seq[IP_CT_DIR_REPLY],
2071                                      cda[CTA_SEQ_ADJ_REPLY]);
2072                 if (ret < 0)
2073                         goto err;
2074
2075                 set_bit(IPS_SEQ_ADJUST_BIT, &ct->status);
2076         }
2077
2078         spin_unlock_bh(&ct->lock);
2079         return 0;
2080 err:
2081         spin_unlock_bh(&ct->lock);
2082         return ret;
2083 }
2084
2085 static const struct nla_policy synproxy_policy[CTA_SYNPROXY_MAX + 1] = {
2086         [CTA_SYNPROXY_ISN]      = { .type = NLA_U32 },
2087         [CTA_SYNPROXY_ITS]      = { .type = NLA_U32 },
2088         [CTA_SYNPROXY_TSOFF]    = { .type = NLA_U32 },
2089 };
2090
2091 static int ctnetlink_change_synproxy(struct nf_conn *ct,
2092                                      const struct nlattr * const cda[])
2093 {
2094         struct nf_conn_synproxy *synproxy = nfct_synproxy(ct);
2095         struct nlattr *tb[CTA_SYNPROXY_MAX + 1];
2096         int err;
2097
2098         if (!synproxy)
2099                 return 0;
2100
2101         err = nla_parse_nested_deprecated(tb, CTA_SYNPROXY_MAX,
2102                                           cda[CTA_SYNPROXY], synproxy_policy,
2103                                           NULL);
2104         if (err < 0)
2105                 return err;
2106
2107         if (!tb[CTA_SYNPROXY_ISN] ||
2108             !tb[CTA_SYNPROXY_ITS] ||
2109             !tb[CTA_SYNPROXY_TSOFF])
2110                 return -EINVAL;
2111
2112         synproxy->isn = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ISN]));
2113         synproxy->its = ntohl(nla_get_be32(tb[CTA_SYNPROXY_ITS]));
2114         synproxy->tsoff = ntohl(nla_get_be32(tb[CTA_SYNPROXY_TSOFF]));
2115
2116         return 0;
2117 }
2118
2119 static int
2120 ctnetlink_attach_labels(struct nf_conn *ct, const struct nlattr * const cda[])
2121 {
2122 #ifdef CONFIG_NF_CONNTRACK_LABELS
2123         size_t len = nla_len(cda[CTA_LABELS]);
2124         const void *mask = cda[CTA_LABELS_MASK];
2125
2126         if (len & (sizeof(u32)-1)) /* must be multiple of u32 */
2127                 return -EINVAL;
2128
2129         if (mask) {
2130                 if (nla_len(cda[CTA_LABELS_MASK]) == 0 ||
2131                     nla_len(cda[CTA_LABELS_MASK]) != len)
2132                         return -EINVAL;
2133                 mask = nla_data(cda[CTA_LABELS_MASK]);
2134         }
2135
2136         len /= sizeof(u32);
2137
2138         return nf_connlabels_replace(ct, nla_data(cda[CTA_LABELS]), mask, len);
2139 #else
2140         return -EOPNOTSUPP;
2141 #endif
2142 }
2143
2144 static int
2145 ctnetlink_change_conntrack(struct nf_conn *ct,
2146                            const struct nlattr * const cda[])
2147 {
2148         int err;
2149
2150         /* only allow NAT changes and master assignation for new conntracks */
2151         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
2152                 return -EOPNOTSUPP;
2153
2154         if (cda[CTA_HELP]) {
2155                 err = ctnetlink_change_helper(ct, cda);
2156                 if (err < 0)
2157                         return err;
2158         }
2159
2160         if (cda[CTA_TIMEOUT]) {
2161                 err = ctnetlink_change_timeout(ct, cda);
2162                 if (err < 0)
2163                         return err;
2164         }
2165
2166         if (cda[CTA_STATUS]) {
2167                 err = ctnetlink_change_status(ct, cda);
2168                 if (err < 0)
2169                         return err;
2170         }
2171
2172         if (cda[CTA_PROTOINFO]) {
2173                 err = ctnetlink_change_protoinfo(ct, cda);
2174                 if (err < 0)
2175                         return err;
2176         }
2177
2178 #if defined(CONFIG_NF_CONNTRACK_MARK)
2179         if (cda[CTA_MARK])
2180                 ctnetlink_change_mark(ct, cda);
2181 #endif
2182
2183         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2184                 err = ctnetlink_change_seq_adj(ct, cda);
2185                 if (err < 0)
2186                         return err;
2187         }
2188
2189         if (cda[CTA_SYNPROXY]) {
2190                 err = ctnetlink_change_synproxy(ct, cda);
2191                 if (err < 0)
2192                         return err;
2193         }
2194
2195         if (cda[CTA_LABELS]) {
2196                 err = ctnetlink_attach_labels(ct, cda);
2197                 if (err < 0)
2198                         return err;
2199         }
2200
2201         return 0;
2202 }
2203
2204 static struct nf_conn *
2205 ctnetlink_create_conntrack(struct net *net,
2206                            const struct nf_conntrack_zone *zone,
2207                            const struct nlattr * const cda[],
2208                            struct nf_conntrack_tuple *otuple,
2209                            struct nf_conntrack_tuple *rtuple,
2210                            u8 u3)
2211 {
2212         struct nf_conn *ct;
2213         int err = -EINVAL;
2214         struct nf_conntrack_helper *helper;
2215         struct nf_conn_tstamp *tstamp;
2216         u64 timeout;
2217
2218         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
2219         if (IS_ERR(ct))
2220                 return ERR_PTR(-ENOMEM);
2221
2222         if (!cda[CTA_TIMEOUT])
2223                 goto err1;
2224
2225         timeout = (u64)ntohl(nla_get_be32(cda[CTA_TIMEOUT])) * HZ;
2226         if (timeout > INT_MAX)
2227                 timeout = INT_MAX;
2228         ct->timeout = (u32)timeout + nfct_time_stamp;
2229
2230         rcu_read_lock();
2231         if (cda[CTA_HELP]) {
2232                 char *helpname = NULL;
2233                 struct nlattr *helpinfo = NULL;
2234
2235                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
2236                 if (err < 0)
2237                         goto err2;
2238
2239                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2240                                                     nf_ct_protonum(ct));
2241                 if (helper == NULL) {
2242                         rcu_read_unlock();
2243 #ifdef CONFIG_MODULES
2244                         if (request_module("nfct-helper-%s", helpname) < 0) {
2245                                 err = -EOPNOTSUPP;
2246                                 goto err1;
2247                         }
2248
2249                         rcu_read_lock();
2250                         helper = __nf_conntrack_helper_find(helpname,
2251                                                             nf_ct_l3num(ct),
2252                                                             nf_ct_protonum(ct));
2253                         if (helper) {
2254                                 err = -EAGAIN;
2255                                 goto err2;
2256                         }
2257                         rcu_read_unlock();
2258 #endif
2259                         err = -EOPNOTSUPP;
2260                         goto err1;
2261                 } else {
2262                         struct nf_conn_help *help;
2263
2264                         help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
2265                         if (help == NULL) {
2266                                 err = -ENOMEM;
2267                                 goto err2;
2268                         }
2269                         /* set private helper data if allowed. */
2270                         if (helper->from_nlattr)
2271                                 helper->from_nlattr(helpinfo, ct);
2272
2273                         /* disable helper auto-assignment for this entry */
2274                         ct->status |= IPS_HELPER;
2275                         RCU_INIT_POINTER(help->helper, helper);
2276                 }
2277         } else {
2278                 /* try an implicit helper assignation */
2279                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
2280                 if (err < 0)
2281                         goto err2;
2282         }
2283
2284         err = ctnetlink_setup_nat(ct, cda);
2285         if (err < 0)
2286                 goto err2;
2287
2288         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
2289         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
2290         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
2291         nf_ct_labels_ext_add(ct);
2292         nfct_seqadj_ext_add(ct);
2293         nfct_synproxy_ext_add(ct);
2294
2295         /* we must add conntrack extensions before confirmation. */
2296         ct->status |= IPS_CONFIRMED;
2297
2298         if (cda[CTA_STATUS]) {
2299                 err = ctnetlink_change_status(ct, cda);
2300                 if (err < 0)
2301                         goto err2;
2302         }
2303
2304         if (cda[CTA_SEQ_ADJ_ORIG] || cda[CTA_SEQ_ADJ_REPLY]) {
2305                 err = ctnetlink_change_seq_adj(ct, cda);
2306                 if (err < 0)
2307                         goto err2;
2308         }
2309
2310         memset(&ct->proto, 0, sizeof(ct->proto));
2311         if (cda[CTA_PROTOINFO]) {
2312                 err = ctnetlink_change_protoinfo(ct, cda);
2313                 if (err < 0)
2314                         goto err2;
2315         }
2316
2317         if (cda[CTA_SYNPROXY]) {
2318                 err = ctnetlink_change_synproxy(ct, cda);
2319                 if (err < 0)
2320                         goto err2;
2321         }
2322
2323 #if defined(CONFIG_NF_CONNTRACK_MARK)
2324         if (cda[CTA_MARK])
2325                 ctnetlink_change_mark(ct, cda);
2326 #endif
2327
2328         /* setup master conntrack: this is a confirmed expectation */
2329         if (cda[CTA_TUPLE_MASTER]) {
2330                 struct nf_conntrack_tuple master;
2331                 struct nf_conntrack_tuple_hash *master_h;
2332                 struct nf_conn *master_ct;
2333
2334                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER,
2335                                             u3, NULL);
2336                 if (err < 0)
2337                         goto err2;
2338
2339                 master_h = nf_conntrack_find_get(net, zone, &master);
2340                 if (master_h == NULL) {
2341                         err = -ENOENT;
2342                         goto err2;
2343                 }
2344                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
2345                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
2346                 ct->master = master_ct;
2347         }
2348         tstamp = nf_conn_tstamp_find(ct);
2349         if (tstamp)
2350                 tstamp->start = ktime_get_real_ns();
2351
2352         err = nf_conntrack_hash_check_insert(ct);
2353         if (err < 0)
2354                 goto err3;
2355
2356         rcu_read_unlock();
2357
2358         return ct;
2359
2360 err3:
2361         if (ct->master)
2362                 nf_ct_put(ct->master);
2363 err2:
2364         rcu_read_unlock();
2365 err1:
2366         nf_conntrack_free(ct);
2367         return ERR_PTR(err);
2368 }
2369
2370 static int ctnetlink_new_conntrack(struct net *net, struct sock *ctnl,
2371                                    struct sk_buff *skb,
2372                                    const struct nlmsghdr *nlh,
2373                                    const struct nlattr * const cda[],
2374                                    struct netlink_ext_ack *extack)
2375 {
2376         struct nf_conntrack_tuple otuple, rtuple;
2377         struct nf_conntrack_tuple_hash *h = NULL;
2378         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2379         struct nf_conn *ct;
2380         u_int8_t u3 = nfmsg->nfgen_family;
2381         struct nf_conntrack_zone zone;
2382         int err;
2383
2384         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
2385         if (err < 0)
2386                 return err;
2387
2388         if (cda[CTA_TUPLE_ORIG]) {
2389                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG,
2390                                             u3, &zone);
2391                 if (err < 0)
2392                         return err;
2393         }
2394
2395         if (cda[CTA_TUPLE_REPLY]) {
2396                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY,
2397                                             u3, &zone);
2398                 if (err < 0)
2399                         return err;
2400         }
2401
2402         if (cda[CTA_TUPLE_ORIG])
2403                 h = nf_conntrack_find_get(net, &zone, &otuple);
2404         else if (cda[CTA_TUPLE_REPLY])
2405                 h = nf_conntrack_find_get(net, &zone, &rtuple);
2406
2407         if (h == NULL) {
2408                 err = -ENOENT;
2409                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2410                         enum ip_conntrack_events events;
2411
2412                         if (!cda[CTA_TUPLE_ORIG] || !cda[CTA_TUPLE_REPLY])
2413                                 return -EINVAL;
2414                         if (otuple.dst.protonum != rtuple.dst.protonum)
2415                                 return -EINVAL;
2416
2417                         ct = ctnetlink_create_conntrack(net, &zone, cda, &otuple,
2418                                                         &rtuple, u3);
2419                         if (IS_ERR(ct))
2420                                 return PTR_ERR(ct);
2421
2422                         err = 0;
2423                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
2424                                 events = 1 << IPCT_RELATED;
2425                         else
2426                                 events = 1 << IPCT_NEW;
2427
2428                         if (cda[CTA_LABELS] &&
2429                             ctnetlink_attach_labels(ct, cda) == 0)
2430                                 events |= (1 << IPCT_LABEL);
2431
2432                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2433                                                       (1 << IPCT_ASSURED) |
2434                                                       (1 << IPCT_HELPER) |
2435                                                       (1 << IPCT_PROTOINFO) |
2436                                                       (1 << IPCT_SEQADJ) |
2437                                                       (1 << IPCT_MARK) |
2438                                                       (1 << IPCT_SYNPROXY) |
2439                                                       events,
2440                                                       ct, NETLINK_CB(skb).portid,
2441                                                       nlmsg_report(nlh));
2442                         nf_ct_put(ct);
2443                 }
2444
2445                 return err;
2446         }
2447         /* implicit 'else' */
2448
2449         err = -EEXIST;
2450         ct = nf_ct_tuplehash_to_ctrack(h);
2451         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
2452                 err = ctnetlink_change_conntrack(ct, cda);
2453                 if (err == 0) {
2454                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
2455                                                       (1 << IPCT_ASSURED) |
2456                                                       (1 << IPCT_HELPER) |
2457                                                       (1 << IPCT_LABEL) |
2458                                                       (1 << IPCT_PROTOINFO) |
2459                                                       (1 << IPCT_SEQADJ) |
2460                                                       (1 << IPCT_MARK) |
2461                                                       (1 << IPCT_SYNPROXY),
2462                                                       ct, NETLINK_CB(skb).portid,
2463                                                       nlmsg_report(nlh));
2464                 }
2465         }
2466
2467         nf_ct_put(ct);
2468         return err;
2469 }
2470
2471 static int
2472 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2473                                 __u16 cpu, const struct ip_conntrack_stat *st)
2474 {
2475         struct nlmsghdr *nlh;
2476         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2477
2478         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
2479                               IPCTNL_MSG_CT_GET_STATS_CPU);
2480         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2481                            NFNETLINK_V0, htons(cpu));
2482         if (!nlh)
2483                 goto nlmsg_failure;
2484
2485         if (nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
2486             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
2487             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
2488             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
2489                                 htonl(st->insert_failed)) ||
2490             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
2491             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
2492             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
2493             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
2494                                 htonl(st->search_restart)) ||
2495             nla_put_be32(skb, CTA_STATS_CLASH_RESOLVE,
2496                                 htonl(st->clash_resolve)))
2497                 goto nla_put_failure;
2498
2499         nlmsg_end(skb, nlh);
2500         return skb->len;
2501
2502 nla_put_failure:
2503 nlmsg_failure:
2504         nlmsg_cancel(skb, nlh);
2505         return -1;
2506 }
2507
2508 static int
2509 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2510 {
2511         int cpu;
2512         struct net *net = sock_net(skb->sk);
2513
2514         if (cb->args[0] == nr_cpu_ids)
2515                 return 0;
2516
2517         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2518                 const struct ip_conntrack_stat *st;
2519
2520                 if (!cpu_possible(cpu))
2521                         continue;
2522
2523                 st = per_cpu_ptr(net->ct.stat, cpu);
2524                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
2525                                                     NETLINK_CB(cb->skb).portid,
2526                                                     cb->nlh->nlmsg_seq,
2527                                                     cpu, st) < 0)
2528                                 break;
2529         }
2530         cb->args[0] = cpu;
2531
2532         return skb->len;
2533 }
2534
2535 static int ctnetlink_stat_ct_cpu(struct net *net, struct sock *ctnl,
2536                                  struct sk_buff *skb,
2537                                  const struct nlmsghdr *nlh,
2538                                  const struct nlattr * const cda[],
2539                                  struct netlink_ext_ack *extack)
2540 {
2541         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2542                 struct netlink_dump_control c = {
2543                         .dump = ctnetlink_ct_stat_cpu_dump,
2544                 };
2545                 return netlink_dump_start(ctnl, skb, nlh, &c);
2546         }
2547
2548         return 0;
2549 }
2550
2551 static int
2552 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
2553                             struct net *net)
2554 {
2555         struct nlmsghdr *nlh;
2556         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2557         unsigned int nr_conntracks = atomic_read(&net->ct.count);
2558
2559         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK, IPCTNL_MSG_CT_GET_STATS);
2560         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
2561                            NFNETLINK_V0, 0);
2562         if (!nlh)
2563                 goto nlmsg_failure;
2564
2565         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
2566                 goto nla_put_failure;
2567
2568         if (nla_put_be32(skb, CTA_STATS_GLOBAL_MAX_ENTRIES, htonl(nf_conntrack_max)))
2569                 goto nla_put_failure;
2570
2571         nlmsg_end(skb, nlh);
2572         return skb->len;
2573
2574 nla_put_failure:
2575 nlmsg_failure:
2576         nlmsg_cancel(skb, nlh);
2577         return -1;
2578 }
2579
2580 static int ctnetlink_stat_ct(struct net *net, struct sock *ctnl,
2581                              struct sk_buff *skb, const struct nlmsghdr *nlh,
2582                              const struct nlattr * const cda[],
2583                              struct netlink_ext_ack *extack)
2584 {
2585         struct sk_buff *skb2;
2586         int err;
2587
2588         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2589         if (skb2 == NULL)
2590                 return -ENOMEM;
2591
2592         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
2593                                           nlh->nlmsg_seq,
2594                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
2595                                           sock_net(skb->sk));
2596         if (err <= 0)
2597                 goto free;
2598
2599         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2600         if (err < 0)
2601                 goto out;
2602
2603         return 0;
2604
2605 free:
2606         kfree_skb(skb2);
2607 out:
2608         /* this avoids a loop in nfnetlink. */
2609         return err == -EAGAIN ? -ENOBUFS : err;
2610 }
2611
2612 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2613         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2614         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2615         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2616         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2617         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2618         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING,
2619                                     .len = NF_CT_HELPER_NAME_LEN - 1 },
2620         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2621         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2622         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2623         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2624         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2625 };
2626
2627 static struct nf_conntrack_expect *
2628 ctnetlink_alloc_expect(const struct nlattr *const cda[], struct nf_conn *ct,
2629                        struct nf_conntrack_helper *helper,
2630                        struct nf_conntrack_tuple *tuple,
2631                        struct nf_conntrack_tuple *mask);
2632
2633 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
2634 static size_t
2635 ctnetlink_glue_build_size(const struct nf_conn *ct)
2636 {
2637         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
2638                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
2639                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
2640                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
2641                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
2642                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
2643                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
2644                + nla_total_size(0) /* CTA_PROTOINFO */
2645                + nla_total_size(0) /* CTA_HELP */
2646                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
2647                + ctnetlink_secctx_size(ct)
2648 #if IS_ENABLED(CONFIG_NF_NAT)
2649                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
2650                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
2651 #endif
2652 #ifdef CONFIG_NF_CONNTRACK_MARK
2653                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
2654 #endif
2655 #ifdef CONFIG_NF_CONNTRACK_ZONES
2656                + nla_total_size(sizeof(u_int16_t)) /* CTA_ZONE|CTA_TUPLE_ZONE */
2657 #endif
2658                + ctnetlink_proto_size(ct)
2659                ;
2660 }
2661
2662 static struct nf_conn *ctnetlink_glue_get_ct(const struct sk_buff *skb,
2663                                              enum ip_conntrack_info *ctinfo)
2664 {
2665         return nf_ct_get(skb, ctinfo);
2666 }
2667
2668 static int __ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct)
2669 {
2670         const struct nf_conntrack_zone *zone;
2671         struct nlattr *nest_parms;
2672
2673         zone = nf_ct_zone(ct);
2674
2675         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG);
2676         if (!nest_parms)
2677                 goto nla_put_failure;
2678         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
2679                 goto nla_put_failure;
2680         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2681                                    NF_CT_ZONE_DIR_ORIG) < 0)
2682                 goto nla_put_failure;
2683         nla_nest_end(skb, nest_parms);
2684
2685         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY);
2686         if (!nest_parms)
2687                 goto nla_put_failure;
2688         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
2689                 goto nla_put_failure;
2690         if (ctnetlink_dump_zone_id(skb, CTA_TUPLE_ZONE, zone,
2691                                    NF_CT_ZONE_DIR_REPL) < 0)
2692                 goto nla_put_failure;
2693         nla_nest_end(skb, nest_parms);
2694
2695         if (ctnetlink_dump_zone_id(skb, CTA_ZONE, zone,
2696                                    NF_CT_DEFAULT_ZONE_DIR) < 0)
2697                 goto nla_put_failure;
2698
2699         if (ctnetlink_dump_id(skb, ct) < 0)
2700                 goto nla_put_failure;
2701
2702         if (ctnetlink_dump_status(skb, ct) < 0)
2703                 goto nla_put_failure;
2704
2705         if (ctnetlink_dump_timeout(skb, ct) < 0)
2706                 goto nla_put_failure;
2707
2708         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
2709                 goto nla_put_failure;
2710
2711         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
2712                 goto nla_put_failure;
2713
2714 #ifdef CONFIG_NF_CONNTRACK_SECMARK
2715         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
2716                 goto nla_put_failure;
2717 #endif
2718         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
2719                 goto nla_put_failure;
2720
2721         if ((ct->status & IPS_SEQ_ADJUST) &&
2722             ctnetlink_dump_ct_seq_adj(skb, ct) < 0)
2723                 goto nla_put_failure;
2724
2725         if (ctnetlink_dump_ct_synproxy(skb, ct) < 0)
2726                 goto nla_put_failure;
2727
2728 #ifdef CONFIG_NF_CONNTRACK_MARK
2729         if (ctnetlink_dump_mark(skb, ct, true) < 0)
2730                 goto nla_put_failure;
2731 #endif
2732         if (ctnetlink_dump_labels(skb, ct) < 0)
2733                 goto nla_put_failure;
2734         return 0;
2735
2736 nla_put_failure:
2737         return -ENOSPC;
2738 }
2739
2740 static int
2741 ctnetlink_glue_build(struct sk_buff *skb, struct nf_conn *ct,
2742                      enum ip_conntrack_info ctinfo,
2743                      u_int16_t ct_attr, u_int16_t ct_info_attr)
2744 {
2745         struct nlattr *nest_parms;
2746
2747         nest_parms = nla_nest_start(skb, ct_attr);
2748         if (!nest_parms)
2749                 goto nla_put_failure;
2750
2751         if (__ctnetlink_glue_build(skb, ct) < 0)
2752                 goto nla_put_failure;
2753
2754         nla_nest_end(skb, nest_parms);
2755
2756         if (nla_put_be32(skb, ct_info_attr, htonl(ctinfo)))
2757                 goto nla_put_failure;
2758
2759         return 0;
2760
2761 nla_put_failure:
2762         return -ENOSPC;
2763 }
2764
2765 static int
2766 ctnetlink_update_status(struct nf_conn *ct, const struct nlattr * const cda[])
2767 {
2768         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
2769         unsigned long d = ct->status ^ status;
2770
2771         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2772                 /* SEEN_REPLY bit can only be set */
2773                 return -EBUSY;
2774
2775         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2776                 /* ASSURED bit can only be set */
2777                 return -EBUSY;
2778
2779         /* This check is less strict than ctnetlink_change_status()
2780          * because callers often flip IPS_EXPECTED bits when sending
2781          * an NFQA_CT attribute to the kernel.  So ignore the
2782          * unchangeable bits but do not error out. Also user programs
2783          * are allowed to clear the bits that they are allowed to change.
2784          */
2785         __ctnetlink_change_status(ct, status, ~status);
2786         return 0;
2787 }
2788
2789 static int
2790 ctnetlink_glue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
2791 {
2792         int err;
2793
2794         if (cda[CTA_TIMEOUT]) {
2795                 err = ctnetlink_change_timeout(ct, cda);
2796                 if (err < 0)
2797                         return err;
2798         }
2799         if (cda[CTA_STATUS]) {
2800                 err = ctnetlink_update_status(ct, cda);
2801                 if (err < 0)
2802                         return err;
2803         }
2804         if (cda[CTA_HELP]) {
2805                 err = ctnetlink_change_helper(ct, cda);
2806                 if (err < 0)
2807                         return err;
2808         }
2809         if (cda[CTA_LABELS]) {
2810                 err = ctnetlink_attach_labels(ct, cda);
2811                 if (err < 0)
2812                         return err;
2813         }
2814 #if defined(CONFIG_NF_CONNTRACK_MARK)
2815         if (cda[CTA_MARK]) {
2816                 ctnetlink_change_mark(ct, cda);
2817         }
2818 #endif
2819         return 0;
2820 }
2821
2822 static int
2823 ctnetlink_glue_parse(const struct nlattr *attr, struct nf_conn *ct)
2824 {
2825         struct nlattr *cda[CTA_MAX+1];
2826         int ret;
2827
2828         ret = nla_parse_nested_deprecated(cda, CTA_MAX, attr, ct_nla_policy,
2829                                           NULL);
2830         if (ret < 0)
2831                 return ret;
2832
2833         return ctnetlink_glue_parse_ct((const struct nlattr **)cda, ct);
2834 }
2835
2836 static int ctnetlink_glue_exp_parse(const struct nlattr * const *cda,
2837                                     const struct nf_conn *ct,
2838                                     struct nf_conntrack_tuple *tuple,
2839                                     struct nf_conntrack_tuple *mask)
2840 {
2841         int err;
2842
2843         err = ctnetlink_parse_tuple(cda, tuple, CTA_EXPECT_TUPLE,
2844                                     nf_ct_l3num(ct), NULL);
2845         if (err < 0)
2846                 return err;
2847
2848         return ctnetlink_parse_tuple(cda, mask, CTA_EXPECT_MASK,
2849                                      nf_ct_l3num(ct), NULL);
2850 }
2851
2852 static int
2853 ctnetlink_glue_attach_expect(const struct nlattr *attr, struct nf_conn *ct,
2854                              u32 portid, u32 report)
2855 {
2856         struct nlattr *cda[CTA_EXPECT_MAX+1];
2857         struct nf_conntrack_tuple tuple, mask;
2858         struct nf_conntrack_helper *helper = NULL;
2859         struct nf_conntrack_expect *exp;
2860         int err;
2861
2862         err = nla_parse_nested_deprecated(cda, CTA_EXPECT_MAX, attr,
2863                                           exp_nla_policy, NULL);
2864         if (err < 0)
2865                 return err;
2866
2867         err = ctnetlink_glue_exp_parse((const struct nlattr * const *)cda,
2868                                        ct, &tuple, &mask);
2869         if (err < 0)
2870                 return err;
2871
2872         if (cda[CTA_EXPECT_HELP_NAME]) {
2873                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2874
2875                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2876                                                     nf_ct_protonum(ct));
2877                 if (helper == NULL)
2878                         return -EOPNOTSUPP;
2879         }
2880
2881         exp = ctnetlink_alloc_expect((const struct nlattr * const *)cda, ct,
2882                                      helper, &tuple, &mask);
2883         if (IS_ERR(exp))
2884                 return PTR_ERR(exp);
2885
2886         err = nf_ct_expect_related_report(exp, portid, report, 0);
2887         nf_ct_expect_put(exp);
2888         return err;
2889 }
2890
2891 static void ctnetlink_glue_seqadj(struct sk_buff *skb, struct nf_conn *ct,
2892                                   enum ip_conntrack_info ctinfo, int diff)
2893 {
2894         if (!(ct->status & IPS_NAT_MASK))
2895                 return;
2896
2897         nf_ct_tcp_seqadj_set(skb, ct, ctinfo, diff);
2898 }
2899
2900 static struct nfnl_ct_hook ctnetlink_glue_hook = {
2901         .get_ct         = ctnetlink_glue_get_ct,
2902         .build_size     = ctnetlink_glue_build_size,
2903         .build          = ctnetlink_glue_build,
2904         .parse          = ctnetlink_glue_parse,
2905         .attach_expect  = ctnetlink_glue_attach_expect,
2906         .seq_adjust     = ctnetlink_glue_seqadj,
2907 };
2908 #endif /* CONFIG_NETFILTER_NETLINK_GLUE_CT */
2909
2910 /***********************************************************************
2911  * EXPECT
2912  ***********************************************************************/
2913
2914 static int ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2915                                     const struct nf_conntrack_tuple *tuple,
2916                                     u32 type)
2917 {
2918         struct nlattr *nest_parms;
2919
2920         nest_parms = nla_nest_start(skb, type);
2921         if (!nest_parms)
2922                 goto nla_put_failure;
2923         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2924                 goto nla_put_failure;
2925         nla_nest_end(skb, nest_parms);
2926
2927         return 0;
2928
2929 nla_put_failure:
2930         return -1;
2931 }
2932
2933 static int ctnetlink_exp_dump_mask(struct sk_buff *skb,
2934                                    const struct nf_conntrack_tuple *tuple,
2935                                    const struct nf_conntrack_tuple_mask *mask)
2936 {
2937         const struct nf_conntrack_l4proto *l4proto;
2938         struct nf_conntrack_tuple m;
2939         struct nlattr *nest_parms;
2940         int ret;
2941
2942         memset(&m, 0xFF, sizeof(m));
2943         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2944         m.src.u.all = mask->src.u.all;
2945         m.src.l3num = tuple->src.l3num;
2946         m.dst.protonum = tuple->dst.protonum;
2947
2948         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK);
2949         if (!nest_parms)
2950                 goto nla_put_failure;
2951
2952         rcu_read_lock();
2953         ret = ctnetlink_dump_tuples_ip(skb, &m);
2954         if (ret >= 0) {
2955                 l4proto = nf_ct_l4proto_find(tuple->dst.protonum);
2956                 ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2957         }
2958         rcu_read_unlock();
2959
2960         if (unlikely(ret < 0))
2961                 goto nla_put_failure;
2962
2963         nla_nest_end(skb, nest_parms);
2964
2965         return 0;
2966
2967 nla_put_failure:
2968         return -1;
2969 }
2970
2971 #if IS_ENABLED(CONFIG_NF_NAT)
2972 static const union nf_inet_addr any_addr;
2973 #endif
2974
2975 static __be32 nf_expect_get_id(const struct nf_conntrack_expect *exp)
2976 {
2977         static __read_mostly siphash_key_t exp_id_seed;
2978         unsigned long a, b, c, d;
2979
2980         net_get_random_once(&exp_id_seed, sizeof(exp_id_seed));
2981
2982         a = (unsigned long)exp;
2983         b = (unsigned long)exp->helper;
2984         c = (unsigned long)exp->master;
2985         d = (unsigned long)siphash(&exp->tuple, sizeof(exp->tuple), &exp_id_seed);
2986
2987 #ifdef CONFIG_64BIT
2988         return (__force __be32)siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &exp_id_seed);
2989 #else
2990         return (__force __be32)siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &exp_id_seed);
2991 #endif
2992 }
2993
2994 static int
2995 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2996                           const struct nf_conntrack_expect *exp)
2997 {
2998         struct nf_conn *master = exp->master;
2999         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
3000         struct nf_conn_help *help;
3001 #if IS_ENABLED(CONFIG_NF_NAT)
3002         struct nlattr *nest_parms;
3003         struct nf_conntrack_tuple nat_tuple = {};
3004 #endif
3005         struct nf_ct_helper_expectfn *expfn;
3006
3007         if (timeout < 0)
3008                 timeout = 0;
3009
3010         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
3011                 goto nla_put_failure;
3012         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
3013                 goto nla_put_failure;
3014         if (ctnetlink_exp_dump_tuple(skb,
3015                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
3016                                  CTA_EXPECT_MASTER) < 0)
3017                 goto nla_put_failure;
3018
3019 #if IS_ENABLED(CONFIG_NF_NAT)
3020         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
3021             exp->saved_proto.all) {
3022                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT);
3023                 if (!nest_parms)
3024                         goto nla_put_failure;
3025
3026                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
3027                         goto nla_put_failure;
3028
3029                 nat_tuple.src.l3num = nf_ct_l3num(master);
3030                 nat_tuple.src.u3 = exp->saved_addr;
3031                 nat_tuple.dst.protonum = nf_ct_protonum(master);
3032                 nat_tuple.src.u = exp->saved_proto;
3033
3034                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
3035                                                 CTA_EXPECT_NAT_TUPLE) < 0)
3036                         goto nla_put_failure;
3037                 nla_nest_end(skb, nest_parms);
3038         }
3039 #endif
3040         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
3041             nla_put_be32(skb, CTA_EXPECT_ID, nf_expect_get_id(exp)) ||
3042             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
3043             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
3044                 goto nla_put_failure;
3045         help = nfct_help(master);
3046         if (help) {
3047                 struct nf_conntrack_helper *helper;
3048
3049                 helper = rcu_dereference(help->helper);
3050                 if (helper &&
3051                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
3052                         goto nla_put_failure;
3053         }
3054         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
3055         if (expfn != NULL &&
3056             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
3057                 goto nla_put_failure;
3058
3059         return 0;
3060
3061 nla_put_failure:
3062         return -1;
3063 }
3064
3065 static int
3066 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
3067                         int event, const struct nf_conntrack_expect *exp)
3068 {
3069         struct nlmsghdr *nlh;
3070         unsigned int flags = portid ? NLM_F_MULTI : 0;
3071
3072         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, event);
3073         nlh = nfnl_msg_put(skb, portid, seq, event, flags,
3074                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3075         if (!nlh)
3076                 goto nlmsg_failure;
3077
3078         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3079                 goto nla_put_failure;
3080
3081         nlmsg_end(skb, nlh);
3082         return skb->len;
3083
3084 nlmsg_failure:
3085 nla_put_failure:
3086         nlmsg_cancel(skb, nlh);
3087         return -1;
3088 }
3089
3090 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3091 static int
3092 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
3093 {
3094         struct nf_conntrack_expect *exp = item->exp;
3095         struct net *net = nf_ct_exp_net(exp);
3096         struct nlmsghdr *nlh;
3097         struct sk_buff *skb;
3098         unsigned int type, group;
3099         int flags = 0;
3100
3101         if (events & (1 << IPEXP_DESTROY)) {
3102                 type = IPCTNL_MSG_EXP_DELETE;
3103                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
3104         } else if (events & (1 << IPEXP_NEW)) {
3105                 type = IPCTNL_MSG_EXP_NEW;
3106                 flags = NLM_F_CREATE|NLM_F_EXCL;
3107                 group = NFNLGRP_CONNTRACK_EXP_NEW;
3108         } else
3109                 return 0;
3110
3111         if (!item->report && !nfnetlink_has_listeners(net, group))
3112                 return 0;
3113
3114         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
3115         if (skb == NULL)
3116                 goto errout;
3117
3118         type = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK_EXP, type);
3119         nlh = nfnl_msg_put(skb, item->portid, 0, type, flags,
3120                            exp->tuple.src.l3num, NFNETLINK_V0, 0);
3121         if (!nlh)
3122                 goto nlmsg_failure;
3123
3124         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
3125                 goto nla_put_failure;
3126
3127         nlmsg_end(skb, nlh);
3128         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
3129         return 0;
3130
3131 nla_put_failure:
3132         nlmsg_cancel(skb, nlh);
3133 nlmsg_failure:
3134         kfree_skb(skb);
3135 errout:
3136         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
3137         return 0;
3138 }
3139 #endif
3140 static int ctnetlink_exp_done(struct netlink_callback *cb)
3141 {
3142         if (cb->args[1])
3143                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
3144         return 0;
3145 }
3146
3147 static int
3148 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3149 {
3150         struct net *net = sock_net(skb->sk);
3151         struct nf_conntrack_expect *exp, *last;
3152         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3153         u_int8_t l3proto = nfmsg->nfgen_family;
3154
3155         rcu_read_lock();
3156         last = (struct nf_conntrack_expect *)cb->args[1];
3157         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
3158 restart:
3159                 hlist_for_each_entry_rcu(exp, &nf_ct_expect_hash[cb->args[0]],
3160                                          hnode) {
3161                         if (l3proto && exp->tuple.src.l3num != l3proto)
3162                                 continue;
3163
3164                         if (!net_eq(nf_ct_net(exp->master), net))
3165                                 continue;
3166
3167                         if (cb->args[1]) {
3168                                 if (exp != last)
3169                                         continue;
3170                                 cb->args[1] = 0;
3171                         }
3172                         if (ctnetlink_exp_fill_info(skb,
3173                                                     NETLINK_CB(cb->skb).portid,
3174                                                     cb->nlh->nlmsg_seq,
3175                                                     IPCTNL_MSG_EXP_NEW,
3176                                                     exp) < 0) {
3177                                 if (!refcount_inc_not_zero(&exp->use))
3178                                         continue;
3179                                 cb->args[1] = (unsigned long)exp;
3180                                 goto out;
3181                         }
3182                 }
3183                 if (cb->args[1]) {
3184                         cb->args[1] = 0;
3185                         goto restart;
3186                 }
3187         }
3188 out:
3189         rcu_read_unlock();
3190         if (last)
3191                 nf_ct_expect_put(last);
3192
3193         return skb->len;
3194 }
3195
3196 static int
3197 ctnetlink_exp_ct_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
3198 {
3199         struct nf_conntrack_expect *exp, *last;
3200         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
3201         struct nf_conn *ct = cb->data;
3202         struct nf_conn_help *help = nfct_help(ct);
3203         u_int8_t l3proto = nfmsg->nfgen_family;
3204
3205         if (cb->args[0])
3206                 return 0;
3207
3208         rcu_read_lock();
3209         last = (struct nf_conntrack_expect *)cb->args[1];
3210 restart:
3211         hlist_for_each_entry_rcu(exp, &help->expectations, lnode) {
3212                 if (l3proto && exp->tuple.src.l3num != l3proto)
3213                         continue;
3214                 if (cb->args[1]) {
3215                         if (exp != last)
3216                                 continue;
3217                         cb->args[1] = 0;
3218                 }
3219                 if (ctnetlink_exp_fill_info(skb, NETLINK_CB(cb->skb).portid,
3220                                             cb->nlh->nlmsg_seq,
3221                                             IPCTNL_MSG_EXP_NEW,
3222                                             exp) < 0) {
3223                         if (!refcount_inc_not_zero(&exp->use))
3224                                 continue;
3225                         cb->args[1] = (unsigned long)exp;
3226                         goto out;
3227                 }
3228         }
3229         if (cb->args[1]) {
3230                 cb->args[1] = 0;
3231                 goto restart;
3232         }
3233         cb->args[0] = 1;
3234 out:
3235         rcu_read_unlock();
3236         if (last)
3237                 nf_ct_expect_put(last);
3238
3239         return skb->len;
3240 }
3241
3242 static int ctnetlink_dump_exp_ct(struct net *net, struct sock *ctnl,
3243                                  struct sk_buff *skb,
3244                                  const struct nlmsghdr *nlh,
3245                                  const struct nlattr * const cda[],
3246                                  struct netlink_ext_ack *extack)
3247 {
3248         int err;
3249         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3250         u_int8_t u3 = nfmsg->nfgen_family;
3251         struct nf_conntrack_tuple tuple;
3252         struct nf_conntrack_tuple_hash *h;
3253         struct nf_conn *ct;
3254         struct nf_conntrack_zone zone;
3255         struct netlink_dump_control c = {
3256                 .dump = ctnetlink_exp_ct_dump_table,
3257                 .done = ctnetlink_exp_done,
3258         };
3259
3260         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3261                                     u3, NULL);
3262         if (err < 0)
3263                 return err;
3264
3265         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3266         if (err < 0)
3267                 return err;
3268
3269         h = nf_conntrack_find_get(net, &zone, &tuple);
3270         if (!h)
3271                 return -ENOENT;
3272
3273         ct = nf_ct_tuplehash_to_ctrack(h);
3274         /* No expectation linked to this connection tracking. */
3275         if (!nfct_help(ct)) {
3276                 nf_ct_put(ct);
3277                 return 0;
3278         }
3279
3280         c.data = ct;
3281
3282         err = netlink_dump_start(ctnl, skb, nlh, &c);
3283         nf_ct_put(ct);
3284
3285         return err;
3286 }
3287
3288 static int ctnetlink_get_expect(struct net *net, struct sock *ctnl,
3289                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3290                                 const struct nlattr * const cda[],
3291                                 struct netlink_ext_ack *extack)
3292 {
3293         struct nf_conntrack_tuple tuple;
3294         struct nf_conntrack_expect *exp;
3295         struct sk_buff *skb2;
3296         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3297         u_int8_t u3 = nfmsg->nfgen_family;
3298         struct nf_conntrack_zone zone;
3299         int err;
3300
3301         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3302                 if (cda[CTA_EXPECT_MASTER])
3303                         return ctnetlink_dump_exp_ct(net, ctnl, skb, nlh, cda,
3304                                                      extack);
3305                 else {
3306                         struct netlink_dump_control c = {
3307                                 .dump = ctnetlink_exp_dump_table,
3308                                 .done = ctnetlink_exp_done,
3309                         };
3310                         return netlink_dump_start(ctnl, skb, nlh, &c);
3311                 }
3312         }
3313
3314         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3315         if (err < 0)
3316                 return err;
3317
3318         if (cda[CTA_EXPECT_TUPLE])
3319                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3320                                             u3, NULL);
3321         else if (cda[CTA_EXPECT_MASTER])
3322                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER,
3323                                             u3, NULL);
3324         else
3325                 return -EINVAL;
3326
3327         if (err < 0)
3328                 return err;
3329
3330         exp = nf_ct_expect_find_get(net, &zone, &tuple);
3331         if (!exp)
3332                 return -ENOENT;
3333
3334         if (cda[CTA_EXPECT_ID]) {
3335                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3336
3337                 if (id != nf_expect_get_id(exp)) {
3338                         nf_ct_expect_put(exp);
3339                         return -ENOENT;
3340                 }
3341         }
3342
3343         err = -ENOMEM;
3344         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
3345         if (skb2 == NULL) {
3346                 nf_ct_expect_put(exp);
3347                 goto out;
3348         }
3349
3350         rcu_read_lock();
3351         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
3352                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
3353         rcu_read_unlock();
3354         nf_ct_expect_put(exp);
3355         if (err <= 0)
3356                 goto free;
3357
3358         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
3359         if (err < 0)
3360                 goto out;
3361
3362         return 0;
3363
3364 free:
3365         kfree_skb(skb2);
3366 out:
3367         /* this avoids a loop in nfnetlink. */
3368         return err == -EAGAIN ? -ENOBUFS : err;
3369 }
3370
3371 static bool expect_iter_name(struct nf_conntrack_expect *exp, void *data)
3372 {
3373         const struct nf_conn_help *m_help;
3374         const char *name = data;
3375
3376         m_help = nfct_help(exp->master);
3377
3378         return strcmp(m_help->helper->name, name) == 0;
3379 }
3380
3381 static bool expect_iter_all(struct nf_conntrack_expect *exp, void *data)
3382 {
3383         return true;
3384 }
3385
3386 static int ctnetlink_del_expect(struct net *net, struct sock *ctnl,
3387                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3388                                 const struct nlattr * const cda[],
3389                                 struct netlink_ext_ack *extack)
3390 {
3391         struct nf_conntrack_expect *exp;
3392         struct nf_conntrack_tuple tuple;
3393         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3394         u_int8_t u3 = nfmsg->nfgen_family;
3395         struct nf_conntrack_zone zone;
3396         int err;
3397
3398         if (cda[CTA_EXPECT_TUPLE]) {
3399                 /* delete a single expect by tuple */
3400                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3401                 if (err < 0)
3402                         return err;
3403
3404                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3405                                             u3, NULL);
3406                 if (err < 0)
3407                         return err;
3408
3409                 /* bump usage count to 2 */
3410                 exp = nf_ct_expect_find_get(net, &zone, &tuple);
3411                 if (!exp)
3412                         return -ENOENT;
3413
3414                 if (cda[CTA_EXPECT_ID]) {
3415                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
3416                         if (ntohl(id) != (u32)(unsigned long)exp) {
3417                                 nf_ct_expect_put(exp);
3418                                 return -ENOENT;
3419                         }
3420                 }
3421
3422                 /* after list removal, usage count == 1 */
3423                 spin_lock_bh(&nf_conntrack_expect_lock);
3424                 if (del_timer(&exp->timeout)) {
3425                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
3426                                                    nlmsg_report(nlh));
3427                         nf_ct_expect_put(exp);
3428                 }
3429                 spin_unlock_bh(&nf_conntrack_expect_lock);
3430                 /* have to put what we 'get' above.
3431                  * after this line usage count == 0 */
3432                 nf_ct_expect_put(exp);
3433         } else if (cda[CTA_EXPECT_HELP_NAME]) {
3434                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3435
3436                 nf_ct_expect_iterate_net(net, expect_iter_name, name,
3437                                          NETLINK_CB(skb).portid,
3438                                          nlmsg_report(nlh));
3439         } else {
3440                 /* This basically means we have to flush everything*/
3441                 nf_ct_expect_iterate_net(net, expect_iter_all, NULL,
3442                                          NETLINK_CB(skb).portid,
3443                                          nlmsg_report(nlh));
3444         }
3445
3446         return 0;
3447 }
3448 static int
3449 ctnetlink_change_expect(struct nf_conntrack_expect *x,
3450                         const struct nlattr * const cda[])
3451 {
3452         if (cda[CTA_EXPECT_TIMEOUT]) {
3453                 if (!del_timer(&x->timeout))
3454                         return -ETIME;
3455
3456                 x->timeout.expires = jiffies +
3457                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
3458                 add_timer(&x->timeout);
3459         }
3460         return 0;
3461 }
3462
3463 #if IS_ENABLED(CONFIG_NF_NAT)
3464 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
3465         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
3466         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
3467 };
3468 #endif
3469
3470 static int
3471 ctnetlink_parse_expect_nat(const struct nlattr *attr,
3472                            struct nf_conntrack_expect *exp,
3473                            u_int8_t u3)
3474 {
3475 #if IS_ENABLED(CONFIG_NF_NAT)
3476         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
3477         struct nf_conntrack_tuple nat_tuple = {};
3478         int err;
3479
3480         err = nla_parse_nested_deprecated(tb, CTA_EXPECT_NAT_MAX, attr,
3481                                           exp_nat_nla_policy, NULL);
3482         if (err < 0)
3483                 return err;
3484
3485         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
3486                 return -EINVAL;
3487
3488         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
3489                                     &nat_tuple, CTA_EXPECT_NAT_TUPLE,
3490                                     u3, NULL);
3491         if (err < 0)
3492                 return err;
3493
3494         exp->saved_addr = nat_tuple.src.u3;
3495         exp->saved_proto = nat_tuple.src.u;
3496         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
3497
3498         return 0;
3499 #else
3500         return -EOPNOTSUPP;
3501 #endif
3502 }
3503
3504 static struct nf_conntrack_expect *
3505 ctnetlink_alloc_expect(const struct nlattr * const cda[], struct nf_conn *ct,
3506                        struct nf_conntrack_helper *helper,
3507                        struct nf_conntrack_tuple *tuple,
3508                        struct nf_conntrack_tuple *mask)
3509 {
3510         u_int32_t class = 0;
3511         struct nf_conntrack_expect *exp;
3512         struct nf_conn_help *help;
3513         int err;
3514
3515         help = nfct_help(ct);
3516         if (!help)
3517                 return ERR_PTR(-EOPNOTSUPP);
3518
3519         if (cda[CTA_EXPECT_CLASS] && helper) {
3520                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
3521                 if (class > helper->expect_class_max)
3522                         return ERR_PTR(-EINVAL);
3523         }
3524         exp = nf_ct_expect_alloc(ct);
3525         if (!exp)
3526                 return ERR_PTR(-ENOMEM);
3527
3528         if (cda[CTA_EXPECT_FLAGS]) {
3529                 exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
3530                 exp->flags &= ~NF_CT_EXPECT_USERSPACE;
3531         } else {
3532                 exp->flags = 0;
3533         }
3534         if (cda[CTA_EXPECT_FN]) {
3535                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
3536                 struct nf_ct_helper_expectfn *expfn;
3537
3538                 expfn = nf_ct_helper_expectfn_find_by_name(name);
3539                 if (expfn == NULL) {
3540                         err = -EINVAL;
3541                         goto err_out;
3542                 }
3543                 exp->expectfn = expfn->expectfn;
3544         } else
3545                 exp->expectfn = NULL;
3546
3547         exp->class = class;
3548         exp->master = ct;
3549         exp->helper = helper;
3550         exp->tuple = *tuple;
3551         exp->mask.src.u3 = mask->src.u3;
3552         exp->mask.src.u.all = mask->src.u.all;
3553
3554         if (cda[CTA_EXPECT_NAT]) {
3555                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
3556                                                  exp, nf_ct_l3num(ct));
3557                 if (err < 0)
3558                         goto err_out;
3559         }
3560         return exp;
3561 err_out:
3562         nf_ct_expect_put(exp);
3563         return ERR_PTR(err);
3564 }
3565
3566 static int
3567 ctnetlink_create_expect(struct net *net,
3568                         const struct nf_conntrack_zone *zone,
3569                         const struct nlattr * const cda[],
3570                         u_int8_t u3, u32 portid, int report)
3571 {
3572         struct nf_conntrack_tuple tuple, mask, master_tuple;
3573         struct nf_conntrack_tuple_hash *h = NULL;
3574         struct nf_conntrack_helper *helper = NULL;
3575         struct nf_conntrack_expect *exp;
3576         struct nf_conn *ct;
3577         int err;
3578
3579         /* caller guarantees that those three CTA_EXPECT_* exist */
3580         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3581                                     u3, NULL);
3582         if (err < 0)
3583                 return err;
3584         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK,
3585                                     u3, NULL);
3586         if (err < 0)
3587                 return err;
3588         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER,
3589                                     u3, NULL);
3590         if (err < 0)
3591                 return err;
3592
3593         /* Look for master conntrack of this expectation */
3594         h = nf_conntrack_find_get(net, zone, &master_tuple);
3595         if (!h)
3596                 return -ENOENT;
3597         ct = nf_ct_tuplehash_to_ctrack(h);
3598
3599         rcu_read_lock();
3600         if (cda[CTA_EXPECT_HELP_NAME]) {
3601                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
3602
3603                 helper = __nf_conntrack_helper_find(helpname, u3,
3604                                                     nf_ct_protonum(ct));
3605                 if (helper == NULL) {
3606                         rcu_read_unlock();
3607 #ifdef CONFIG_MODULES
3608                         if (request_module("nfct-helper-%s", helpname) < 0) {
3609                                 err = -EOPNOTSUPP;
3610                                 goto err_ct;
3611                         }
3612                         rcu_read_lock();
3613                         helper = __nf_conntrack_helper_find(helpname, u3,
3614                                                             nf_ct_protonum(ct));
3615                         if (helper) {
3616                                 err = -EAGAIN;
3617                                 goto err_rcu;
3618                         }
3619                         rcu_read_unlock();
3620 #endif
3621                         err = -EOPNOTSUPP;
3622                         goto err_ct;
3623                 }
3624         }
3625
3626         exp = ctnetlink_alloc_expect(cda, ct, helper, &tuple, &mask);
3627         if (IS_ERR(exp)) {
3628                 err = PTR_ERR(exp);
3629                 goto err_rcu;
3630         }
3631
3632         err = nf_ct_expect_related_report(exp, portid, report, 0);
3633         nf_ct_expect_put(exp);
3634 err_rcu:
3635         rcu_read_unlock();
3636 err_ct:
3637         nf_ct_put(ct);
3638         return err;
3639 }
3640
3641 static int ctnetlink_new_expect(struct net *net, struct sock *ctnl,
3642                                 struct sk_buff *skb, const struct nlmsghdr *nlh,
3643                                 const struct nlattr * const cda[],
3644                                 struct netlink_ext_ack *extack)
3645 {
3646         struct nf_conntrack_tuple tuple;
3647         struct nf_conntrack_expect *exp;
3648         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
3649         u_int8_t u3 = nfmsg->nfgen_family;
3650         struct nf_conntrack_zone zone;
3651         int err;
3652
3653         if (!cda[CTA_EXPECT_TUPLE]
3654             || !cda[CTA_EXPECT_MASK]
3655             || !cda[CTA_EXPECT_MASTER])
3656                 return -EINVAL;
3657
3658         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
3659         if (err < 0)
3660                 return err;
3661
3662         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE,
3663                                     u3, NULL);
3664         if (err < 0)
3665                 return err;
3666
3667         spin_lock_bh(&nf_conntrack_expect_lock);
3668         exp = __nf_ct_expect_find(net, &zone, &tuple);
3669         if (!exp) {
3670                 spin_unlock_bh(&nf_conntrack_expect_lock);
3671                 err = -ENOENT;
3672                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
3673                         err = ctnetlink_create_expect(net, &zone, cda, u3,
3674                                                       NETLINK_CB(skb).portid,
3675                                                       nlmsg_report(nlh));
3676                 }
3677                 return err;
3678         }
3679
3680         err = -EEXIST;
3681         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
3682                 err = ctnetlink_change_expect(exp, cda);
3683         spin_unlock_bh(&nf_conntrack_expect_lock);
3684
3685         return err;
3686 }
3687
3688 static int
3689 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
3690                              const struct ip_conntrack_stat *st)
3691 {
3692         struct nlmsghdr *nlh;
3693         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
3694
3695         event = nfnl_msg_type(NFNL_SUBSYS_CTNETLINK,
3696                               IPCTNL_MSG_EXP_GET_STATS_CPU);
3697         nlh = nfnl_msg_put(skb, portid, seq, event, flags, AF_UNSPEC,
3698                            NFNETLINK_V0, htons(cpu));
3699         if (!nlh)
3700                 goto nlmsg_failure;
3701
3702         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
3703             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
3704             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
3705                 goto nla_put_failure;
3706
3707         nlmsg_end(skb, nlh);
3708         return skb->len;
3709
3710 nla_put_failure:
3711 nlmsg_failure:
3712         nlmsg_cancel(skb, nlh);
3713         return -1;
3714 }
3715
3716 static int
3717 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
3718 {
3719         int cpu;
3720         struct net *net = sock_net(skb->sk);
3721
3722         if (cb->args[0] == nr_cpu_ids)
3723                 return 0;
3724
3725         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
3726                 const struct ip_conntrack_stat *st;
3727
3728                 if (!cpu_possible(cpu))
3729                         continue;
3730
3731                 st = per_cpu_ptr(net->ct.stat, cpu);
3732                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
3733                                                  cb->nlh->nlmsg_seq,
3734                                                  cpu, st) < 0)
3735                         break;
3736         }
3737         cb->args[0] = cpu;
3738
3739         return skb->len;
3740 }
3741
3742 static int ctnetlink_stat_exp_cpu(struct net *net, struct sock *ctnl,
3743                                   struct sk_buff *skb,
3744                                   const struct nlmsghdr *nlh,
3745                                   const struct nlattr * const cda[],
3746                                   struct netlink_ext_ack *extack)
3747 {
3748         if (nlh->nlmsg_flags & NLM_F_DUMP) {
3749                 struct netlink_dump_control c = {
3750                         .dump = ctnetlink_exp_stat_cpu_dump,
3751                 };
3752                 return netlink_dump_start(ctnl, skb, nlh, &c);
3753         }
3754
3755         return 0;
3756 }
3757
3758 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3759 static struct nf_ct_event_notifier ctnl_notifier = {
3760         .fcn = ctnetlink_conntrack_event,
3761 };
3762
3763 static struct nf_exp_event_notifier ctnl_notifier_exp = {
3764         .fcn = ctnetlink_expect_event,
3765 };
3766 #endif
3767
3768 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
3769         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
3770                                             .attr_count = CTA_MAX,
3771                                             .policy = ct_nla_policy },
3772         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
3773                                             .attr_count = CTA_MAX,
3774                                             .policy = ct_nla_policy },
3775         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
3776                                             .attr_count = CTA_MAX,
3777                                             .policy = ct_nla_policy },
3778         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
3779                                             .attr_count = CTA_MAX,
3780                                             .policy = ct_nla_policy },
3781         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
3782         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
3783         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
3784         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
3785 };
3786
3787 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
3788         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
3789                                             .attr_count = CTA_EXPECT_MAX,
3790                                             .policy = exp_nla_policy },
3791         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
3792                                             .attr_count = CTA_EXPECT_MAX,
3793                                             .policy = exp_nla_policy },
3794         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
3795                                             .attr_count = CTA_EXPECT_MAX,
3796                                             .policy = exp_nla_policy },
3797         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
3798 };
3799
3800 static const struct nfnetlink_subsystem ctnl_subsys = {
3801         .name                           = "conntrack",
3802         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
3803         .cb_count                       = IPCTNL_MSG_MAX,
3804         .cb                             = ctnl_cb,
3805 };
3806
3807 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
3808         .name                           = "conntrack_expect",
3809         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
3810         .cb_count                       = IPCTNL_MSG_EXP_MAX,
3811         .cb                             = ctnl_exp_cb,
3812 };
3813
3814 MODULE_ALIAS("ip_conntrack_netlink");
3815 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
3816 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
3817
3818 static int __net_init ctnetlink_net_init(struct net *net)
3819 {
3820 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3821         int ret;
3822
3823         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
3824         if (ret < 0) {
3825                 pr_err("ctnetlink_init: cannot register notifier.\n");
3826                 goto err_out;
3827         }
3828
3829         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
3830         if (ret < 0) {
3831                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
3832                 goto err_unreg_notifier;
3833         }
3834 #endif
3835         return 0;
3836
3837 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3838 err_unreg_notifier:
3839         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3840 err_out:
3841         return ret;
3842 #endif
3843 }
3844
3845 static void ctnetlink_net_exit(struct net *net)
3846 {
3847 #ifdef CONFIG_NF_CONNTRACK_EVENTS
3848         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
3849         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
3850 #endif
3851 }
3852
3853 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
3854 {
3855         struct net *net;
3856
3857         list_for_each_entry(net, net_exit_list, exit_list)
3858                 ctnetlink_net_exit(net);
3859
3860         /* wait for other cpus until they are done with ctnl_notifiers */
3861         synchronize_rcu();
3862 }
3863
3864 static struct pernet_operations ctnetlink_net_ops = {
3865         .init           = ctnetlink_net_init,
3866         .exit_batch     = ctnetlink_net_exit_batch,
3867 };
3868
3869 static int __init ctnetlink_init(void)
3870 {
3871         int ret;
3872
3873         ret = nfnetlink_subsys_register(&ctnl_subsys);
3874         if (ret < 0) {
3875                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
3876                 goto err_out;
3877         }
3878
3879         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
3880         if (ret < 0) {
3881                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
3882                 goto err_unreg_subsys;
3883         }
3884
3885         ret = register_pernet_subsys(&ctnetlink_net_ops);
3886         if (ret < 0) {
3887                 pr_err("ctnetlink_init: cannot register pernet operations\n");
3888                 goto err_unreg_exp_subsys;
3889         }
3890 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3891         /* setup interaction between nf_queue and nf_conntrack_netlink. */
3892         RCU_INIT_POINTER(nfnl_ct_hook, &ctnetlink_glue_hook);
3893 #endif
3894         return 0;
3895
3896 err_unreg_exp_subsys:
3897         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3898 err_unreg_subsys:
3899         nfnetlink_subsys_unregister(&ctnl_subsys);
3900 err_out:
3901         return ret;
3902 }
3903
3904 static void __exit ctnetlink_exit(void)
3905 {
3906         unregister_pernet_subsys(&ctnetlink_net_ops);
3907         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
3908         nfnetlink_subsys_unregister(&ctnl_subsys);
3909 #ifdef CONFIG_NETFILTER_NETLINK_GLUE_CT
3910         RCU_INIT_POINTER(nfnl_ct_hook, NULL);
3911 #endif
3912         synchronize_rcu();
3913 }
3914
3915 module_init(ctnetlink_init);
3916 module_exit(ctnetlink_exit);